CSPC-BRS: An Enhanced Real-Time Multi-Target Detection and Tracking Algorithm for Complex Open Channels

Abstract

Ensuring worker safety compliance and secure cargo transportation in complex port environments is critical for modern logistics hubs. However, conventional supervision methods, including manual inspection and passive video monitoring, suffer from limited coverage, poor real-time responsiveness, and low robustness under frequent occlusion, scale variation, and cross-camera transitions, leading to unstable target association and missed risk events. To address these challenges, this paper proposes CSPC-BRS, a real-time multi-object detection and tracking framework for open-channel port scenarios. CSPC (Coordinated Spatial Perception Cascade) enhances the YOLOv8 backbone by integrating CASAM, SPPELAN-DW, and CACC modules to improve feature representation under cluttered backgrounds and degraded visual conditions. Meanwhile, BRS (Bounding Box Reduction Strategy) mitigates scale distortion during tracking, and a Multi-Dimensional Re-identification Scoring (MDRS) mechanism fuses six perceptual features—color, texture, shape, motion, size, and time—to achieve stable cross-camera identity consistency. Experimental results demonstrate that CSPC-BRS outperforms the YOLOv8-n baseline by improving the mAP@0.5:0.95 by 9.6% while achieving a real-time speed of 132.63 FPS. Furthermore, in practical deployment, it reduces the false capture rate by an average of 59.7% compared to the YOLOv8 + Bot-SORT tracker. These results confirm that CSPC-BRS effectively balances detection accuracy and computational efficiency, providing a practical and deployable solution for intelligent safety monitoring in complex industrial logistics environments.

1. Introduction

In real-world industrial environments, such as port terminals and logistics hubs, ensuring proper use of personal protective equipment by workers and compliance of vehicles with operational regulations is essential for maintaining production safety and operational efficiency [1]. However, existing safety supervision methods still rely heavily on manual inspections or passive video review, which tend to be labor-intensive, limited in coverage, and often insufficient for achieving real-time and continuous monitoring. These limitations become more pronounced in complex open-channel scenarios characterized by challenging illumination, frequent occlusions, and dynamic interactions between personnel and vehicles. Under such conditions, human supervision may struggle to consistently and promptly detect safety violations (e.g., unauthorized e-bikes, workers without helmets), thereby motivating the development of intelligent algorithms capable of performing accurate and real-time multi-object detection and tracking (MOT) to improve the automation and reliability of industrial safety management.

The computer vision community has achieved substantial progress in object detection and tracking, largely driven by deep learning. Pioneering detection frameworks like Faster R-CNN [2], SSD [3], and the YOLO series [4] have established strong baselines for performance and speed. For this work, we select YOLOv8 [5] as our baseline, motivated by its exceptional balance of accuracy, efficiency, and extensive community support, which facilitates replication and deployment. The core contributions of this paper—the proposed CSPC modules—are architectural and can be integrated into newer detector backbones; our choice of YOLOv8 serves as a robust and widely recognized platform for validation. Concurrently, attention mechanisms have emerged as a powerful tool to enhance feature representation. As surveyed by Guo et al. [6], these mechanisms, including channel and spatial attention, dynamically modulate feature weights, significantly boosting performance in tasks like image classification and object detection. Techniques such as the Convolutional Block Attention Module (CBAM) [7] and subsequent variants have been successfully integrated into deep networks to suppress irrelevant features and focus on critical regions. Beyond standard visual modalities, research into multi-modal fusion, such as RGB-Thermal salient object detection [8], highlights advanced strategies for robust feature integration under challenging conditions, offering valuable insights for complex industrial settings.

In the pursuit of real-time performance, several studies have focused on optimizing model architecture. For instance, Cho and Kim’s Selective Attention Network (SANet) [9] demonstrated robust detection in complex industrial settings, while Zhou’s YOLO-NL [10] incorporated a non-local attention mechanism to improve multi-scale detection. To further reduce computational overhead, depthwise separable convolutions [11] have been widely adopted to create more efficient network designs without substantial performance loss, a principle also effectively employed in lightweight detectors like NanoDet [12].

In the domain of multi-object tracking, methods like DeepSORT [13] and ByteTrack [14] have become benchmarks. DeepSORT enhances the SORT algorithm [15] by integrating appearance descriptors via a deep association metric, improving long-term tracking. ByteTrack demonstrates the importance of associating every detection box, including low-score ones, to reduce identity switches. Other approaches, such as the fusion of 2D LIDAR and camera data [16] or the combination of enhanced kernelized correlation filters with YOLO [17], highlight the trend towards multi-modal and hybrid tracking solutions for improved robustness.

Despite these advancements, several notable challenges remain when deploying these technologies in complex open industrial environments, such as port bridge scenarios. First, although attention mechanisms can enhance feature extraction, their integration frequently increases computational complexity, which may adversely affect real-time performance [18], particularly in edge deployment scenarios where computational resources are limited. Second, existing detection models continue to face difficulties in achieving robust multi-scale feature fusion under conditions of severe occlusion, variable lighting, and low contrast between targets and the background [19]. Third, in tracking tasks, frequent occlusions and parallel or intersecting motion of multiple targets may result in increased occurrences of target loss and identity (ID) switching [20]. Methods such as ByteTrack demonstrate promising improvements, yet their performance may degrade when targets experience significant perspective changes or prolonged occlusion. Finally, achieving consistent identity association across multiple camera views remains challenging in large-scale and complex environments, as most existing tracking algorithms are primarily designed for single-camera scenarios.

To address these multifaceted challenges, this study proposes CSPC-BRS, a task-oriented framework that integrates the Coordinated Spatial Perception Cascade (CSPC) detection component with the Bounding Box Reduction Strategy (BRS). This framework combines an enhanced YOLOv8-based detector with an optimized Bot-SORT tracker, and is specifically tailored to address the complexities of open-channel surveillance environments. While the current study validates the framework on a standard GPU platform to establish its performance baseline, its design incorporates efficiency-oriented principles that are conducive to future edge deployment. The main contributions of this study can be summarized as follows:

(1)

An enhanced detection architecture based on the Coordinated Spatial Perception Cascade (CSPC): The CSPC is constructed by integrating three key modules into the YOLOv8 framework: a Channel and Spatial Attention Module (CASAM) in the backbone, a Spatial Pyramid Pooling Efficient Layer Aggregation Network with Depthwise Convolution (SPPELAN-DW) replacing the original SPPF, and Convolutional Attention Channel Combination (CACC) modules in the neck network. This cascade enhances multi-scale feature extraction and fusion under challenging conditions while maintaining a relatively high inference speed.

(2)

A tracking strategy with bounding box stabilization: An optimized Bot-SORT algorithm is incorporated and further enhanced by a Bounding Box Reduction Strategy (BRS). This strategy dynamically adjusts bounding box dimensions based on motion prediction and historical information, contributing to the mitigation of target loss and ID switches for fast-moving and intersecting objects.

(3)

A multi-dimensional re-identification mechanism: A Multi-Dimensional Re-Identification Scoring (MDRS) mechanism is proposed, which fuses six key attributes—color, texture, shape, motion, size, and temporal consistency—into a weighted similarity score, contributing to improved ID consistency across different viewpoints and under occlusion.

2. Approaches

2.1. Processes and Frameworks

To address the challenges of identity switching, scale distortion, and environmental interference in cross-camera tracking within river-port logistics environments, this paper proposes the CSPC-BRS framework, whose overall architecture is illustrated in Figure 1. The framework integrates three core components: an enhanced detection module, a multi-dimensional re-identification module, and an adaptive tracking module. The process begins with the CSPC detection module, which incorporates three key sub-modules—CASAM (Channel and Spatial Attention Module), SPPELAN-DW (Spatial Pyramid Pooling Efficient Layer Aggregation Network with Depthwise Convolution), and CACC (Convolutional Attention Channel Combination)—to achieve robust feature extraction and target localization under challenging conditions such as reflection, low contrast, and adverse weather. The detected objects are then forwarded to the BRS tracking module, which employs a bounding-box rescaling strategy to dynamically adjust for size variations caused by viewpoint changes, thereby maintaining consistent target proportions across different camera views. Simultaneously, the MDRS (Multi-Dimensional Re-Identification Scoring) module computes a similarity score by fusing six feature dimensions: color, texture, shape, motion, size, and time. This score is used for cross-camera identity association, ensuring reliable ID preservation even under occlusion or perspective shift. By synergistically combining detection, re-identification, and tracking, CSPC-BRS achieves high-confidence, low-identity-switch multi-camera object tracking, providing a dependable technological foundation for intelligent safety management in port logistics systems.

2.2. Coordinated Spatial Perception Cascade (CSPC)

YOLO-based detectors have become a dominant paradigm in real-time object detection due to their favorable balance between accuracy and efficiency. Among them, YOLOv8, released by Ultralytics in 2023, provides a mature and stable baseline architecture for industrial deployment, particularly in dynamic monitoring environments that demand consistent performance and reproducible optimization. However, when directly applied to open-channel port scenes, conventional YOLOv8 exhibits notable limitations, including sensitivity to reflective interference, insufficient robustness to extreme scale variation, and unstable feature fusion across hierarchical layers.

To overcome these deficiencies, we construct the Coordinated Spatial Perception Cascade (CSPC), which integrates three problem-oriented enhancement modules into the YOLOv8 framework: Channel and Spatial Attention Module (CASAM) and SPPELAN-DW embedded within the backbone, along with the Convolutional Attention Channel Combination (CACC) deployed in the neck stage. Rather than solely increasing architectural complexity, CSPC is designed as a coordinated perception refinement mechanism, where each module targets a specific failure mode commonly observed in complex port surveillance scenarios.

Specifically, CASAM focuses on mitigating channel dominance caused by reflective surfaces and low-contrast backgrounds, SPPELAN-DW addresses scale inconsistency and receptive field imbalance arising from variable camera viewpoints, while CACC improves semantic consistency during multi-scale feature fusion. This structured integration strategy, illustrated in Figure 2, enhances feature stability and discriminative representation, enabling the detector to maintain reliable performance under fluctuating illumination, dense occlusion, and heterogeneous object distributions.

Overall, CSPC is not designed as a structural expansion of YOLOv8, but as a coordinated refinement framework that systematically addresses stability, scale inconsistency, and semantic misalignment specific to open-channel surveillance environments, thereby forming a targeted perception optimization cascade.

Figure 2 illustrates the coordinated information flow among these modules, emphasizing their complementary roles in stabilizing feature extraction and reducing redundant activation under industrial operating conditions.

2.2.1. CASAM Attention Mechanism

The CASAM is a problem-oriented attention mechanism specifically designed to enhance the feature representation capability of the YOLOv8 backbone under complex open-channel conditions. In our experimental scenario, reflective interference primarily originates from safety helmet surfaces rather than metallic backgrounds, combined with fluctuating illumination, motion blur, and frequent partial occlusions. These factors often lead to unstable channel responses, local overexposure, and spatial ambiguity around helmet edges and contours. To address these issues, CASAM integrates channel attention, spatial attention, self-attention, and a channel reordering strategy into a coordinated refinement process that suppresses redundant activations and strengthens discriminative feature patterns specific to safety helmet detection.

The overall architecture of CASAM is illustrated in Figure 3, where the input feature map is first subjected to channel reordering, followed by parallel attention computation paths that collaboratively regulate both channel-wise and spatial feature distributions.

The channel attention component operates globally to model inter-channel dependencies and identify discriminative channels. Given an input feature map x, global average pooling (GAP) is first used to compress spatial information into a channel descriptor, which is subsequently transformed through two fully connected layers to generate channel attention weights, as expressed in Equation (1):

$$\mathrm{C}\mathrm{A}\left(x\right)=\sigma \left(W_{2}ReLU\left(W_{1}GAP\left(x\right)\right)\right)$$

(1)

where x denotes the input feature map, GAP represents global average pooling, W₁ and W₂ are the weights of the two fully connected layers responsible for channel dimension reduction and restoration, respectively, ReLU is the Rectified Linear Unit activation function, and σ is the Sigmoid function that maps the output to the range [0, 1]. These weights adaptively emphasize informative channels while suppressing less relevant ones. Compared to conventional channel attention mechanisms such as CBAM, this formulation provides more stable channel discrimination under reflective helmet interference, glare, and non-uniform illumination [21].

To capture long-range spatial dependencies and internal structural relationships, CASAM introduces a self-attention mechanism, enabling the model to correlate spatially distant but semantically related regions. This process is formulated as Equation (2):

$$\mathrm{Attention}\left(Q,K,V\right)=softmax\left({\displaystyle \frac{Q{K}^T}{\sqrt{d_k}}}\right)$$

(2)

where Q, K, and V represent the query, key, and value matrices derived from linear projections of the input feature map, QK^T computes the similarity between feature vectors, d_k is the dimensionality of the key vectors, and the scaling factor √d_k stabilizes gradient flow during training. This mechanism enhances the model’s ability to recognize safety helmet contours and structural continuity that may be disrupted by glare, specular highlights, or background clutter.

The spatial attention component focuses on highlighting informative spatial regions and suppressing background noise. It generates a spatial attention map by aggregating average-pooling and max-pooling results across the channel dimension, followed by convolutional fusion, as described in Equation (3):

$$\mathrm{S}\mathrm{A}\left(x\right)=\sigma \left(\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}\left(\left[AvgPool\left(x\right),MaxPool\left(x\right)\right]\right)\right)$$

(3)

where AvgPool and MaxPool denote average and maximum pooling operations along the channel axis, respectively. The resulting feature maps are concatenated and processed by a convolution layer, and the Sigmoid function produces spatial attention coefficients. This mechanism improves localization precision and strengthens sensitivity to salient helmet regions, especially under cluttered backgrounds and visually confusing conditions.

To further enhance feature robustness, CASAM incorporates a structured and differentiable channel reordering operation prior to attention computation. This process performs controlled permutation of channel groups, effectively disrupting fixed inter-channel dependencies and preventing feature co-adaptation. By exposing the network to dynamically reconfigured channel arrangements, this strategy encourages the learning of more generalized and disentangled feature representations, thereby improving robustness against reflective distortions caused by safety helmet surfaces and diverse environmental conditions.

The introduction of Channel Reordering prior to attention computation serves as a structured perturbation and regularization mechanism during training. In standard training, convolutional filters may develop fixed, co-adaptive dependencies among feature channels specific to the training data distribution. This co-adaptation can lead to fragile representations that over-rely on specific channel constellations, reducing robustness to novel variations (e.g., unseen reflective patterns on safety helmets). By performing a differentiable, controlled permutation of channel groups before the attention module computes its weights, we explicitly disrupt these static inter-channel correlations. This forces the attention mechanism to learn more generalizable and disentangled feature relationships, as it cannot rely on a fixed channel ordering. Consequently, the model develops a more robust attentional focus that is invariant to specific channel arrangements, which is particularly beneficial for stabilizing feature responses under the variable lighting and reflective interference prevalent in open-channel surveillance.

This integration strategy, together with the channel reordering mechanism, establishes CASAM as a targeted enhancement module optimized for safety helmet detection in high-variability port surveillance scenarios, rather than a simple aggregation of existing attention techniques.

2.2.2. SPPELAN-DW Aggregation Network

The SPPELAN-DW module, illustrated in Figure 4, is introduced as a scale-stabilized aggregation structure to replace the conventional SPPF module in YOLOv8. Its design specifically targets the pronounced scale variation observed in open-channel surveillance, where safety helmets and workers can appear at drastically different sizes due to camera distance, perspective changes, and dynamic movement. While the standard SPPF module expands the receptive field effectively, it often produces redundant feature responses and lacks explicit control over scale imbalance, which may weaken the representation of small or distant targets.

To overcome these limitations, SPPELAN-DW integrates Spatial Pyramid Pooling (SPP) with Depthwise Separable Convolution (DWConv) within an ELAN-based aggregation framework [22], forming a lightweight yet scale-aware feature refinement pathway. The input feature map is first processed through depthwise convolution layers, which decouple spatial and channel-wise computations. This operation reduces parameter redundancy and computational complexity while preserving critical structural information, such as the contour stability and spatial integrity of safety helmet regions.

Subsequently, a sequence of interconnected convolution layers performs two successive 3 × 3 depthwise convolutions, reinforcing nonlinear feature transformation and improving the stability of spatial responses under heterogeneous target scales. The refined feature map is then forwarded to a multi-branch pooling stage composed of three max-pooling operations with different kernel sizes and identical stride. This parallel configuration enables the module to encode contextual information at multiple receptive field levels, allowing adaptive perception of small, medium, and large target instances without disproportionately amplifying any single scale.

After multi-scale pooling, all pooled feature maps are concatenated along the channel dimension to form a unified multi-scale representation. This representation is further compressed and fused through a 1 × 1 convolution layer, which performs dimensionality reduction and feature integration, preserving the most salient scale-sensitive features while suppressing redundant activations.

In contrast to conventional implementations that directly replace standard convolutions with lightweight operations, SPPELAN-DW is designed as a structural scale regulator that explicitly balances receptive field expansion and computational efficiency. By strategically embedding DWConv throughout the aggregation pathway, the module achieves a controlled trade-off between expressive capacity and real-time performance, leading to improved detection stability for targets exhibiting extreme scale variation, such as distant or partially obscured safety helmets.

This design establishes SPPELAN-DW as a purpose-driven enhancement to YOLOv8, functioning not as a simple architectural replacement but as a dedicated mechanism for scale-aware feature stabilization in safety-critical open-channel surveillance scenarios.

2.2.3. CACC Convolutions

Although the neck network in YOLOv8 effectively integrates multi-scale features, its standard convolutional layers lack an explicit mechanism for adaptive channel recalibration, limiting their ability to dynamically suppress less informative channels and emphasize task-critical ones. This deficiency becomes particularly problematic in cluttered open-channel environments, where subtle target features such as safety helmets are easily confused with background elements, leading to false activations and degraded detection reliability. To address this limitation, we introduce the Convolutional Attention Channel Combination (CACC) module as a dedicated enhancement for the neck network.

CACC is designed as an attention-augmented convolutional unit that tightly couples channel attention with the convolution operation itself, rather than treating attention as an isolated post-processing step. The module begins with a standard convolution layer that transforms the input feature map while preserving spatial structure. This is followed by batch normalization to stabilize feature distribution and improve training convergence, and a nonlinear activation function (SiLU by default) to enhance representational expressiveness.

The core of CACC lies in its channel attention mechanism, which evaluates the relative importance of each channel based on global contextual information [23]. Specifically, global average pooling is applied to generate a channel descriptor vector, which is then passed through two fully connected layers to produce channel-wise attention weights. These weights are subsequently used to reweight the convolutional feature maps through element-wise multiplication, amplifying salient feature responses and suppressing noisy or redundant activations.

Unlike conventional approaches such as SE-Net, where attention is applied after the convolutional operation as an auxiliary recalibration step, CACC embeds channel attention directly within the convolutional block, forming an integrated feature refinement unit. This design enables more coherent interaction between spatial transformation and channel prioritization, resulting in more precise feature discrimination during multi-scale fusion.

Furthermore, the deployment of CACC at the neck stage is a strategic design choice. Positioned immediately before the fusion of feature maps from different scales, CACC functions as a cross-scale feature purifier, selectively filtering out background interference and reinforcing semantically consistent target features across hierarchical layers. This targeted application enhances inter-scale semantic alignment and reduces misactivation caused by scale inconsistency or background clutter.

Through its deep integration of attention and convolution within the neck network, CACC transcends the role of a generic attention module and serves as a specialized mechanism for enhancing feature continuity and discrimination in safety helmet detection tasks. This structural innovation distinguishes CACC from existing attention paradigms and contributes to the overall robustness of the CSPC in complex surveillance environments.

2.3. BRS

In open channels, the dynamic nature of targets—characterized by viewpoint variations, occlusions, and varying velocities—induces significant fluctuations in the apparent size and shape of bounding boxes. These fluctuations are a primary cause of identity switches (ID switches) and tracking failures, particularly when targets intersect or move in parallel. To address this fundamental challenge, we propose a novel Bounding Box Reduction Strategy (BRS). Specifically, BRS unifies a Kalman Filter-based motion predictor with a confidence-weighted update rule for adaptive size stabilization.

2.3.1. Kalman Filter for Motion State Prediction

The BRS employs a Kalman filter to maintain a stable estimate of the target’s kinematic state, which forms the prior for our subsequent bounding box adjustment. We define the state vector at time $k$ as $x_k=[c_x,c_y,v_x,v_y,w,h,w,h{]}^T$, encompassing the bounding box center coordinates $(c_x,c_y)$, their velocities $(v_x,v_y)$, the width $w$ and height $h$, and their rates of change $(w,h)$.

The standard Kalman filter recursion, as given in Equation (4), provides the optimal estimate of this state.

$$\left\{\begin{array}{c}{\widehat{x}}_{k|k-1}=F\cdot {\widehat{x}}_{k-1|k-1}+B\cdot u_k\\ P_{k|k-1}=F\cdot P_{k-1|k-1}\cdot F^T+Q\\ K_k=P_{k|k-1}\cdot H^T\cdot (H\cdot P_{k|k-1}\cdot H^T+R{)}^{-1}\\ {\widehat{x}}_{k|k}={\widehat{x}}_{k|k-1}+K_k\cdot (z_k-H\cdot {\widehat{x}}_{k|k-1})\\ P_{k|k}=(I-K_k\cdot H)\cdot P_{k|k-1}\end{array}\right.$$

(4)

Here, $F$ is the state transition matrix, $Q$ and $R$ are the process and observation noise covariance matrices, $K_k$ is the Kalman gain, and $H$ is the observation matrix. The filter outputs a predicted state ${\widehat{x}}_{k|k-1}$ and a corresponding covariance $P_{k|k-1}$, which quantifies the uncertainty of this prediction.

2.3.2. Bounding Box Refinement: A Simplified Bayesian Framework

The Kalman filter provides a prediction for the entire state, including the bounding box dimensions $(w,h)$. However, the raw detection $z_k$ often provides a noisy and unstable estimate of size, especially under occlusion or perspective distortion, where we seek the posterior distribution of the true dimensions $\theta =[w,h{]}^T$ given all observed data $D$.

We define the components of Bayes’ theorem (Equation (5)) for our specific task:

$$p\left(\theta |D\right)={\displaystyle \frac{p\left(D|\theta \right)\cdot p\left(\theta \right)}{p\left(D\right)}}$$

(5)

Prior $p\left(\theta \right)$: We use the Kalman filter’s predicted dimensions ${\theta }_{\mathrm{pred}}=[w_{\mathrm{pred}},h_{\mathrm{pred}}{]}^T$ from ${\widehat{\mathrm{x}}}_{k\mid k-1}$ to form a Gaussian prior: $p\left(\theta \right)=\mathcal{N}(\theta \mid {\theta }_{\mathrm{pred}},{\mathsf{\Sigma }}_{\mathrm{pred}})$. The covariance ${\mathsf{\Sigma }}_{\mathrm{p}\mathrm{red}}$ is extracted from the relevant elements of $p_{k\mid k-1}$, encapsulating the prediction uncertainty.

Likelihood $p(D\mid \theta )$ This term models the probability of observing the current detector’s output ${\theta }_{\det }=[w_{\det },h_{\det }{]}^T$ given a hypothetical true size $\theta$. We model this as $p(D\mid \theta )=\mathcal{N}({\theta }_{\det }\mid \theta ,{\mathsf{\Sigma }}_{\det })$, where ${\mathsf{\Sigma }}_{\det }$ is a fixed covariance matrix representing the detector’s inherent noise.

Under these Gaussian assumptions, the posterior $p(\theta \mid D)$ is also Gaussian. The mean of this posterior, which is the optimal estimate under a minimum mean squared error criterion, is given by a convex combination of the prior mean (Kalman prediction) and the observation (detector output), weighted by their respective precisions (inverse covariances).

2.3.3. The BRS Update Rule

Solving for the posterior mean under the Gaussian assumptions leads to an optimal estimate that is a weighted sum of the prior (Kalman prediction) and the observation (detector output). This principle can be broadly expressed for a single dimension (e.g., width $w$) as:

$$w_k=(1-\alpha )\cdot w_{k-1}+\alpha \cdot w_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d}}+\beta \cdot (w_{\mathrm{d}\mathrm{e}\mathrm{t}}-w_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d}})$$

(6)

where the coefficients $\alpha$ and $\beta$ are derived from the respective precisions (inverse covariances) of the prior and the likelihood.

To instantiate this principle with high computational efficiency and empirical robustness, we propose a simplified yet effective update rule that operates on the collective bounding box dimensions. The closed-form solution under Gaussian assumptions leads to an update rule that is a weighted sum of the prior and observation. Inspired by this Bayesian filtering paradigm and for real-time computational efficiency, we derive a simplified yet effective instantiation.

This leads to our final BRS formulation, which updates the bounding box size vector $B_k=[w,h{]}_k^T$ as follows:

$$B_k=B_{k-1}\cdot \gamma +\alpha \cdot B_{\mathrm{pred}}$$

(7)

$$\gamma =1-\exp \left(-{\displaystyle \frac{(c_x-p_x{)}^2+(c_y-p_y{)}^2}{2{\sigma }^2}}\right)$$

(8)

In this formulation, the bounding box update is governed by a localization confidence weight $\gamma$ (Equation (8)), which dynamically balances the contribution between the historical size and the motion-predicted size. The weight $\gamma$ is computed from the Euclidean distance between the detected center $(c_x,c_y)$ and the Kalman-predicted center $(p_x,p_y)$, thereby reflecting the instantaneous reliability of the detection. A high-confidence detection with minimal center discrepancy yields $\gamma \approx 1$, strongly anchoring the new estimate $B_k$ to the stable prior size $B_{k-1}$ to ensure temporal smoothness and reduce jitter. Conversely, a low-confidence detection caused by occlusion or noise drives $\gamma \approx 0$, effectively discounting the potentially unreliable prior state and allowing the Kalman prediction $B_{\mathrm{pred}}$ to dominate the update. This design decouples the dependency on historical size from the confidence in the current observation, enabling the bounding box to remain stable during unreliable detections while proactively adapting to predictable scale variations through the term $\alpha \cdot B_{\mathrm{pred}}$. The parameter $\alpha$ controls the overall influence of the motion-predicted size, while $\sigma$ modulates the system’s sensitivity to localization inaccuracies. As a principled approximation of Bayesian filtering, this mechanism directly mitigates the root cause of identity switches—erratic size fluctuations—by ensuring robust size consistency across frames.

2.4. MDRS

To achieve robust cross-camera identity consistency in challenging port environments, we propose a Multi-Dimensional Re-Identification Scoring (MDRS) mechanism. This mechanism integrates six complementary feature dimensions—color, texture, shape, motion, size, and temporal consistency—to compute a robust matching score between object instances across different camera views. We empirically evaluated several candidate weight combinations and selected Configuration A as our final setting, as it achieved the highest IDF1 score on the validation set.

The specific steps include the following steps:

(1)

Color similarity: calculates the cosine similarity of multi-channel color features such as RGB, HSV, and grayscale;

(2)

Texture similarity: extract grayscale texture and edge texture information (such as Sobel edges);

(3)

Shape similarity: use complexity and aspect ratio to model shape consistency;

(4)

Motion direction similarity: the consistency of relative direction is calculated based on the center point velocity vector of the front and back frames;

(5)

Dimensional similarity: Calculate regional scale consistency based on the difference in aspect ratio;

(6)

Time similarity factor: According to the dynamic attenuation effect of time difference, it avoids long-term target mismatch.

This mechanism computes individual similarity scores across each feature dimension and applies a weighted fusion strategy to derive a final confidence score for cross-camera matching. The feature components are extracted from consecutive frames across different camera perspectives, and the matching score is used in the ID association stage of CSPC-BRS.

The final scoring formula is defined as:

$${\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}}_{\mathrm{f}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{l}}=w_c\cdot \mathrm{C}\mathrm{o}\mathrm{l}\mathrm{o}\mathrm{r}+w_t\cdot \mathrm{T}\mathrm{e}\mathrm{x}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{e}+w_s\cdot \mathrm{S}\mathrm{h}\mathrm{a}\mathrm{p}\mathrm{e}+w_m\cdot \mathrm{M}\mathrm{o}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}+w_z\cdot \mathrm{S}\mathrm{i}\mathrm{z}\mathrm{e}+w_{\tau }\cdot \mathrm{T}\mathrm{i}\mathrm{m}\mathrm{e}$$

(9)

where the weights are fixed as $w_c=0.40$, $w_t=0.20$, $w_s=0.15$, $w_m=0.15$, $w_z=0.05$, and $w_{\tau }=0.05$. This formula is implemented in Algorithm 1, which outlines the end-to-end process for computing the multi-dimensional similarity between two target observations.

Pseudo-code of the algorithm is presented in Algorithm 1.

Algorithm 1: Multi-Dimensional Re-ID Similarity Mechanism

Input:

-Feature 1, Feature 2: Feature vectors of two targets

-Motion 1, Motion 2: Velocity vectors

-Size 1, Size 2: Object sizes (area or diagonal length)

-TimeDiff: Frame interval between observations

Output:

-FinalScore: Aggregated similarity score

-ScoreDict: Individual similarity components

1:    Initialize ScoreDict = empty dictionary

2:    if color features exist:

3:            HSVsim = CosineSim (Feature 1. color. HSV, Feature 2.color. HSV)

4:            RGBsim = CosineSim (Feature 1. color. RGB, Feature 2.color. RGB)

5:            Graysim = CosineSim (Feature 1. color. Gray, Feature 2.color. Gray)

6:            ScoreDict [“Color”] = Average(HSVsim, RGBsim, Graysim)

7:    else:

8:            ScoreDict [“Color”] = 0

9: if texture features exist:

10:          Graysim = CosineSim(Feature 1. texture. Gray, Feature 2.texture. Gray)

11:          Edgesim = CosineSim(Feature 1. texture. Edges, Feature 2.texture. Edges)

12:          ScoreDict [“Texture”] = Average(Graysim, Edgesim)

13: else:

14:          ScoreDict [“Texture”] = 0

15: if shape features exist:

16:          Complexity = 1 − |C1 − C2|/max(C1, C2, ε)

17:          Aspect = 1 − |AR1 − AR2|/max(AR1, AR2, ε)

18:          ScoreDict [“Shape”] = Average(Complexity, Aspect)

19: else:

20:          ScoreDict [“Shape”] = 0

21: if motion features exist:

22:          ScoreDict [“Motion”] = VectorSim(Motion 1, Motion 2)

23: else:

24:          ScoreDict [“Motion”] = 0

25: ScoreDict [“Size”] = min(Size 1/Size 2, Size 2/Size 1)

26: TimeFactor = max(0, 1 − TimeDiff/120)

27: ScoreDict [“Time”] = TimeFactor

28: FinalScore =

0.4 * ScoreDict [“Color”]+  // `*` denotes multiplication

0.2 * ScoreDict [“Texture”]+

0.15 * ScoreDict [“Shape”]+

0.15 * ScoreDict [“Motion”]+

0.05 * ScoreDict [“Size”]+

0.05 * ScoreDict [“Time”]

29: return FinalScore, ScoreDict

2.5. Evaluation Metrics

This comprehensive study meticulously evaluates the performance of the network model by employing a diverse array of metrics, each tailored to assess different aspects of the model’s effectiveness. Key among these are recall, precision, and mean Average Precision (mAP@0.5) across multiple categories, offering a nuanced understanding of the model’s ability to accurately classify various objects or entities. Additionally, computational complexity is analyzed to gauge the model’s efficiency, ensuring practicality for real-world applications.

To further underscore the model’s suitability for real-time usage, Frames Per Second (FPS) is utilized as a crucial benchmark. As defined in Equation (10), FPS quantifies the number of frames the model can process per second, taking into account the cumulative time required for pre_processing, inference, and post_processing. In our experiments, the time for each stage is measured in milliseconds (ms). This holistic approach ensures that not only is the model’s accuracy evaluated but also its speed, which is paramount for seamless integration into dynamic environments.

$$\left\{\begin{array}{c}\hfill \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}Precision={\displaystyle \frac{TP}{TP+FP}} \hfill \\ \hfill \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}Recall={\displaystyle \frac{TP}{TP+FN}} \hfill \\ \\ \hfill \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}AP={\displaystyle {\mathit{\int }}_0^1} Precision\left(Recall\right)dRecall\hfill \\ mAP={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N} A{P}_i\\ FPS={\displaystyle \frac{1000}{pr{e}_{\_}process+inference+pos{t}_{\_}process}}\end{array}\right.$$

(10)

Specifically, precision is calculated as the ratio of true positives (TP) to the sum of true positives and false positives (FP), indicating the model’s precision in identifying positive samples. Recall, on the other hand, measures the model’s ability to find all positive samples, expressed as the ratio of true positives to the total number of positive samples (including both true positives and false negatives, FN).

Average Precision (AP) integrates precision across varying recall levels, providing a more nuanced view of the model’s performance. By averaging AP across all categories (mAP), a single score encapsulates the overall performance across multiple tasks, making it a valuable metric for comparing different models.

Finally, the inclusion of FPS as a performance metric underscores the study’s commitment to evaluating the model’s practicality for real-time applications. By considering the total time required for pre_processing, inference, and post_processing, the study ensures that the model’s speed is assessed holistically, accounting for the full range of operations involved in deploying the model in practice.

In summary, this study employs a rigorous and multifaceted approach to evaluating the performance of the network model, ensuring that both accuracy and real-time capability are thoroughly assessed and validated.

3. Results and Analysis

3.1. Dataset

The dataset employed in this study was constructed to address real-world safety monitoring requirements in port logistics environments, with a strong emphasis on task specificity, annotation accuracy, and operational authenticity. Although the total image count (3150) is smaller than that of large-scale generic benchmarks, it represents a highly specialized and densely annotated collection derived from three months of continuous surveillance video streams and historical violation archives provided by a cooperative port authority in Anhui Province, China. The dataset prioritizes ecological validity over sheer volume, ensuring that the data faithfully reflect realistic deployment conditions faced by industrial safety monitoring systems. It should be emphasized that this dataset is not intended as a generic benchmark, but rather as a domain-focused evaluation environment optimized for validating deployment-oriented performance under realistic constraints.

Images were captured from four operational channels under continuous monitoring and encompass a wide range of practical challenges, including pronounced target scale variation, frequent inter-object occlusion, background clutter, reflective helmet surfaces, perspective distortion, and heterogeneous illumination conditions (daylight, dusk, nighttime with artificial lighting, and shadowed regions). These factors introduce significant visual complexity and closely mirror real operational scenarios, thereby enhancing the credibility and practical relevance of performance evaluation in safety-critical applications.

The dataset comprises seven categories relevant to port safety supervision: helmet, head, person, truck, fake_helmet, tricycle, and bicycle. Representative visual examples for each category under various challenges are provided in Figure 5b. All annotations were manually verified to ensure labeling accuracy and consistency. In total, more than 8000 object instances were annotated, resulting in high annotation density and providing a strong supervisory signal for model training. The dataset was partitioned into training, validation, and testing sets using a 7:2:1 ratio, corresponding to 2205 images for training, 630 for validation, and 315 for testing, ensuring reliable and unbiased performance assessment.

Figure 5a provides a statistical overview of the dataset. The class distribution reflects natural operational imbalance, consistent with real-world working conditions. The bounding box size distribution and width–height scatter plots indicate a predominance of small- and medium-scale targets, confirming the dataset’s inherent detection difficulty. Furthermore, the spatial distribution of target centers demonstrates broad coverage across the image plane, reducing positional bias and enhancing spatial diversity.

We openly acknowledge that the dataset does not yet reach the scale of tens of thousands of images nor encompass multiple geographic regions or seasonal contexts. Accordingly, this study positions CSPC-BRS as a domain-focused and deployment-oriented solution evaluated under a domain-consistent setting. Large-scale multi-port, multi-season dataset expansion and cross-domain validation will be essential directions for future work to further quantify the model’s transferability. Nevertheless, the current dataset, characterized by realistic environmental complexity, and statistical diversity, as shown in Figure 5b, provides a rigorous and appropriate foundation for validating the effectiveness and reliability of the proposed method in practical port surveillance scenarios.

3.2. Experimental Environment and Parameter Settings

The experimental environment for this study was configured to ensure fairness, reproducibility, and stability in all comparative evaluations. All experiments were conducted on a workstation equipped with a Windows 10 operating system, an Intel (R) Core (TM) i5-9300H CPU, and an NVIDIA GeForce RTX 1650 GPU (4 GB VRAM). The deep learning framework employed was PyTorch 1.13.0, with CUDA 11.6 and cuDNN 8.4 acceleration enabled. All inference-related performance measurements were executed using batch size = 1 to simulate real-time video stream processing.

For real-time performance evaluation, we adopted a standardized benchmarking protocol:

• 50 warm-up iterations were excluded from timing measurement
• Inference time was averaged over 1000 consecutive forward passes
• All models were tested in FP32 precision mode
• The GPU was exclusively allocated during testing to avoid background interference

The detailed experimental environment configuration is summarized in

Table 1. Experimental environment.

Name	Parameter
System	Windows 10
CPU	Intel(R) Core(TM) i5-9300H
GPU	NVIDIA GeForce RTX 1650
Framework	Pytorch (1.13.0)
CUDA	11.6
cuDNN	8.4
Precision	FP32

.

To ensure optimal convergence and stable training, the parameter settings were carefully selected as shown in

Table 2. Experimental Parameter Settings.

Parameter	Settings
epochs	100
imgsz	640
optimizer	SGD
iou	0.6
momentum	0.937
batch size	8 (training)/1 (inference)

. The training process was conducted for 100 epochs, enabling sufficient learning while minimizing the risk of overfitting. The input image resolution (imgsz) was set to 640 × 640, balancing detection accuracy and computational efficiency.

The Stochastic Gradient Descent (SGD) optimizer was employed due to its robust convergence characteristics, with an IoU threshold of 0.6 to balance localization accuracy and recall performance. A momentum coefficient of 0.937 was adopted to accelerate convergence while reducing optimization oscillations.

3.3. Detection Results and Comprehensive Analysis

The loss function, serving as a pivotal metric, quantifies the discrepancy between the model’s forecasts and the ground truth values, thereby exerting a direct influence on the overall performance of the detection model. As depicted in Figure 6, a discernible trend emerges during the training and validation phases, where the constituent losses—encompassing box_loss, cls_loss, and dfl_loss—undergo a gradual decline and eventually stabilize. This progression indicates stable convergence of the model towards optimal parameters, ensuring that the prediction errors are minimized.

Furthermore, the model’s Precision and Recall metrics, which are crucial indicators of its detection prowess, exhibit remarkable values nearing 0.9 and 0.95, respectively, at the culmination of the training process. These high figures underscore the model’s exceptional ability to accurately identify true positives while minimizing false negatives, thereby demonstrating a robust detection accuracy and recall capability.

To provide a more comprehensive assessment of the model’s performance, we also evaluate it using the mean Average Precision (mAP) metric at various Intersection over Union (IoU) thresholds. Specifically, the model achieves an mAP@0.5 score of approximately 0.95, indicating a strong performance in detecting objects with a high degree of overlap with the ground truth. Moreover, the mAP@0.5:0.95 score of approximately 0.6 demonstrates the model’s resilience and adaptability across a broader range of IoU thresholds, further validating its comprehensive detection capabilities.

In summary, the consistent decline and stabilization of loss functions, coupled with the impressive Precision, Recall, and mAP scores, collectively attest to the effectiveness and robustness of the proposed detection model. These results underscore the model’s accurate detection capability and its practical potential for real-world applications where high detection accuracy and recall are paramount.

A confusion matrix serves as a comprehensive visualization tool for assessing the algorithmic performance, offering insights into how well the model distinguishes between different classes. In this matrix, each row meticulously represents the actual class distribution, whereas each column portrays the model’s predicted class outcomes. Notably, the elements aligned along the main diagonal hold paramount importance, as they signify the number of correctly classified samples, commonly referred to as True Positives (TP).

Delving deeper, the lower-left triangular region of the matrix unveils a crucial aspect of the model’s performance—False Negatives (FN). These represent instances where the model fails to detect actual samples, inadvertently misclassifying them as belonging to other categories or overlooking them entirely. This region highlights the model’s limitations in recognizing certain targets, potentially leading to unrecognized or mislabeled objects.

Conversely, the upper-right triangular region encapsulates False Positives (FP), instances where the model erroneously classifies background elements or samples from different categories as belonging to the current category. A high density of FP values indicates that the model may suffer from a higher rate of spurious detections, mistakenly identifying non-target objects as belonging to the current class.

As evidenced in Figure 7a, the robust presence of elements along the main diagonal underscores the high degree of accuracy achieved by our detection model across most categories. For instance, the ‘helmet’ category has a TP of 323, and the ‘person’ category has a TP of 615, testifying to the model’s exceptional ability to discern these targets from others. Notably, even in the challenging “fake_helmet” category, characterized by its similarity to “head” and “helmet,” the model manages to maintain a relatively low level of both false and missed detections.

Expanding on this analysis, Figure 7b provides a more nuanced view of the model’s performance by presenting the accuracy for each category. The “helmet” category achieves an accuracy of 95%, the “person” category with 91%, and the ‘truck’ category achieves an accuracy of 100% under the given test conditions, indicating high performance for this category. However, a closer inspection of the normalized matrix also reveals nuances in the model’s behavior. Specifically, the “background” category experiences a 0.30 probability of being misclassified as “person” and a 0.27 probability of being mislabeled as “bicycle,” indicating that ambient noise and background complexities in real-world scenarios can still pose challenges to the model’s classification accuracy. Nonetheless, these findings serve as valuable insights for future model refinements and optimizations to further enhance its robustness and performance.

3.3.1. Ablation Experiment

To meticulously assess the individual and collective contributions of our proposed modules—SPPELAN-DW, CASAM, and CACC—to the overall performance of our model, we designed and executed a comprehensive ablation study. By systematically incorporating and excluding these modules within the YOLOv8 framework, we delved into their specific impacts on both detection accuracy, measured by mAP@0.5 and the more stringent mAP@0.5~0.95, and processing speed, quantified by frames per second (FPS). The findings of this rigorous experimentation are summarized in

Table 3. Ablation experiment results.

SPPELAN-DW	CASAM	CACC	mAP@0.5	mAP@0.5~0.95	FPS (ms)
×	×	×	0.922	0.570	123.25
√	×	×	0.931	0.589	129.47
×	√	×	0.934	0.610	127.50
×	×	√	0.932	0.612	126.54
×	√	√	0.938	0.599	142.30
√	×	√	0.936	0.612	146.64
√	√	×	0.935	0.595	144.21
√	√	√	0.949	0.625	132.63

.

Table 3 presents the ablation study results, where each row represents a unique configuration of the modules. The checkmark (√) signifies the inclusion of a module, whereas the cross (×) denotes its exclusion. The standalone integration of SPPELAN-DW not only elevated the mAP@0.5 to 0.931 but also bolstered the FPS to 129.47, underscoring its pivotal role in enhancing both detection precision and computational efficiency. Similarly, the introduction of CASAM alone significantly improved the mAP@0.5~0.95 to 0.610, emphasizing its effectiveness in boosting accuracy across a broader range of IoU thresholds.

The CACC module, when integrated, demonstrated a dual benefit, elevating both the mAP@0.5 to 0.932 and FPS to 126.54, highlighting its ability to harmoniously optimize both accuracy and speed. The synergistic effect of combining CASAM and CACC was evident, with the mAP@0.5 soaring to 0.938 and FPS achieving an impressive 142.30, despite a marginal dip in mAP@0.5~0.95, which remained at a high level.

The pairing of SPPELAN-DW and CACC yielded an mAP@0.5 of 0.936 coupled with an exceptional FPS of 146.64, showcasing their complementary strengths in accelerating processing while preserving accuracy. Likewise, the combination of SPPELAN-DW and CASAM struck a fine balance between accuracy (mAP@0.5 = 0.935) and speed (FPS = 144.21), further validating the efficacy of our proposed modules.

The unified integration of the SPPELAN-DW, CASAM, and CACC modules delivers the best performance, achieving a peak mAP@0.5 of 0.949 and an elevated mAP@0.5~0.95 of 0.625, without compromising real-time performance (132.63 FPS). This outcome highlights the success of our modular design in achieving a favorable speed-accuracy trade-off, where each component works synergistically to boost detection robustness and efficiency in complex open channels.

3.3.2. Comparative Experiment

To rigorously evaluate the detection performance of the proposed CSPC-BRS framework, we conducted comprehensive comparative experiments against several state-of-the-art detectors, including YOLOv5-n, YOLOv6-n, YOLOv7-n, YOLOv8-n, RT-DETR-R18, PP-YOLOE+-s, and NanoDet-Plus. This systematic evaluation aimed to provide a thorough performance analysis across different architectural paradigms.

Figure 8 illustrates the learning trajectories of all compared models throughout the training process. Our CSPC-BRS model demonstrates rapid convergence in mAP@0.5 during the initial training phase. After approximately 20 epochs, it maintains a stable improvement trajectory, consistently outperforming all other methods. In contrast, while the YOLO series models show reasonable convergence trends, they exhibit noticeable performance fluctuations in later training stages. The CSPC-BRS model ultimately achieves the highest mAP@0.5 of 0.949 within 100 epochs, demonstrating superior training stability and convergence behavior.

As presented in

Table 4. Comprehensive Comparison of Detection Performance.

Model	Parameters (M)	GFLOPs	mAP@0.5	mAP@0.5:0.95	Improvement vs. YOLOv8 (mAP@0.5:0.95)
YOLOv5-n [24]	1.7	4.3	0.920	0.557	−2.3%
YOLOv6-n [25]	4.2	11.8	0.922	0.566	−0.7%
YOLOv7-n [26]	6.0	13.2	0.909	0.521	−8.6%
YOLOv8-n (Baseline) [27]	3.0	8.1	0.925	0.570	Reference
RT-DETR-R18 [28]	5.9	13.1	0.915	0.568	−0.4%
PP-YOLOE+-s [29]	3.2	8.9	0.927	0.579	+1.6%
NanoDet-Plus [30]	1.1	0.8	0.865	0.452	−20.7%
CSPC-BRS (Ours)	3.6	8.6	0.949	0.625	+9.6%

, the proposed CSPC-BRS architecture achieves a superior balance between detection accuracy and computational complexity. Compared with the YOLOv8-n baseline, CSPC-BRS introduces a moderate increase in model size, with parameters rising from 3.0 M to 3.6 M (+20%) and GFLOPs from 8.1 to 8.6 (+6.2%). This additional computational cost results in a substantial improvement in detection accuracy, with mAP@0.5:0.95 increasing from 0.570 to 0.625, corresponding to a relative improvement of 9.6%.

The comparative results also indicate that higher computational complexity does not necessarily guarantee superior accuracy. For example, YOLOv7-n, despite having the highest GFLOPs (13.2) and parameters (6.0 M), exhibits the lowest mAP@0.5:0.95 among the YOLO series. In contrast, CSPC-BRS achieves a 19.9% higher accuracy than YOLOv7-n while reducing parameters by 40% and FLOPs by 34.8%, highlighting the effectiveness of its structural optimization.

Table 5. Analysis of Computational Efficiency and Real-Time Performance.

Model	Pre-Processing (ms)	Inference (ms)	Post-Processing (ms)	FPS
YOLOv5-n	0.3	5.1	2.1	131.15
YOLOv6-n	0.5	6.2	3.4	98.92
YOLOv7-n	0.5	7.5	3.3	89.28
YOLOv8-n	0.4	5.6	0.9	145.35
RT-DETR-R18	0.5	8.0	0.6	107.52
PP-YOLOE+-s	0.4	6.3	1.5	121.95
NanoDet-Plus	0.3	2.2	1.8	235.29
CSPC-BRS (Ours)	0.4	6.4	1.4	132.63

summarizes the runtime performance measured under unified experimental conditions. YOLOv8-n achieves an inference time of 5.6 ms and a frame rate of 145.35 FPS, while CSPC-BRS requires 6.4 ms per frame, corresponding to 132.63 FPS. Although CSPC-BRS exhibits slightly lower real-time speed due to its increased computational load, it maintains a high-performance level suitable for real-time industrial deployment while delivering significantly enhanced detection accuracy.

NanoDet-Plus achieves the highest frame rate (235.29 FPS) owing to its extremely lightweight design; however, its substantial performance degradation (−20.7% mAP@0.5:0.95 relative to YOLOv8-n) limits its applicability in safety-critical scenarios where precision is paramount.

Overall, these results demonstrate that CSPC-BRS prioritizes precision-critical detection performance while preserving practical real-time operability. The modest trade-off in speed is justified by its notable gain in detection reliability.

To comprehensively evaluate the robustness and detection performance of the CSPC-BRS framework in complex open channel environments, we selected four representative scenarios (A–D) from the validation set imagery. For visual comparison clarity, we focused on comparing CSPC-BRS against the YOLO series models (YOLOv5-n, YOLOv6-n, YOLOv7-n, and YOLOv8-n), as these architectures share similar design paradigms, enabling more straightforward visual analysis of detection differences.

Figure 9 clearly demonstrates the superior performance of CSPC-BRS across all test scenarios. In scenario (A), characterized by an employee wearing a hat resembling a helmet amidst inclement weather and obstructions, YOLOv5-n and YOLOv6-n failed to detect the “fake_helmet” target entirely. While YOLOv7-n and YOLOv8-n successfully identified the target, both models generated false positives by misclassifying it as a genuine “helmet.” In contrast, CSPC-BRS accurately identified the “fake_helmet” with high confidence, effectively eliminating false positives through its refined feature extraction capabilities and the effective Bounding Box Reduction Strategy (BRS).

Scenario (B) presented a challenging situation with overlapping targets, where a “person” was partially occluded by a “tricycle,” and a small “helmet” blended with the background. In this case, YOLOv5-n and YOLOv6-n exhibited significant missed detections, failing to identify multiple targets. YOLOv7-n and YOLOv8-n demonstrated improved detection capability but with notably lower confidence scores and less precise bounding box localization compared to CSPC-BRS. Our method achieved significantly higher confidence scores and more accurate bounding boxes, demonstrating its exceptional capability in handling occlusion and detecting small objects.

Scenarios (C) and (D), featuring inadequate lighting conditions and significant target size variations, further revealed the limitations of the baseline models. YOLOv5-n and YOLOv6-n consistently showed the poorest performance with numerous missed detections across these challenging conditions. While YOLOv7-n and YOLOv8-n exhibited comparable detection performance to each other, both suffered from inconsistent confidence scores and occasional false positives. The proposed CSPC-BRS, empowered by its multi-level attention mechanism and optimized feature fusion, maintained robust detection performance with high confidence scores and minimal false positives even under low-light conditions and with varying target sizes.

The visual comparative analysis demonstrates that CSPC-BRS achieves more stable and reliable detection performance compared to other models. While YOLOv5-n and YOLOv6-n suffer from substantial missed detections, and YOLOv7-n/YOLOv8-n exhibit inconsistent performance with false positives, our method consistently outperforms all YOLO variants in terms of both detection accuracy and confidence stability, demonstrating its practical advantage for real-world deployment in complex port environments.

3.3.3. Interpretability Analysis via Gradient-Weighted Class Activation Mapping

To obtain a more interpretable understanding of the behavioral divergence between the baseline models and our proposed CSPC-BRS, we utilize Gradient-weighted Class Activation Mapping (Grad-CAM) [31]. This technique generates coarse localization maps by leveraging the gradient information of a specific target class (e.g., helmet) flowing into the final convolutional layer, thereby highlighting image regions most influential for the model’s prediction. This analysis allows us to examine and compare the spatial focus of each detector during the decision-making process.

In this analysis, Grad-CAM was applied to the final convolutional layers of YOLOv5-n, YOLOv6-n, YOLOv7-n, YOLOv8-n, and our CSPC-BRS model. We selected four representative validation images exhibiting challenging conditions including low illumination, strong glare, shadow interference, and the presence of small targets. The resulting class activation heatmaps for all models are comparatively presented in Figure 10.

The comparative visualization yields several observations. For the models YOLOv5-n to YOLOv7-n, the activation is frequently dispersed or misallocated to background clutter or high-contrast edges, particularly under strong glare or complex shadows, which aligns with their observed higher false positive rates and missed detections of small objects such as fake_helmet. While YOLOv8-n demonstrates improved suppression of diffuse background activation compared to earlier versions, its focus on critical small targets (head, helmet, fake_helmet) remains weak, characterized by faint and poorly localized activations. In contrast, the heatmaps for our CSPC-BRS model show more concentrated activation precisely on the small target regions, even under adverse lighting, alongside a marked reduction in activation within non-target areas. This indicates that the integrated CASAM and CACC modules enhance the amplification of discriminative foreground features while suppressing distracting background patterns. These visual observations correlate with the quantitative performance gains reported in Table 4 and Table 5, suggesting that the architectural refinements in CSPC-BRS contribute to more stable and interpretable perceptual focus in complex scenes.

3.3.4. Actual Detection Results

To comprehensively evaluate the robustness and efficacy of our detection model, we conducted a rigorous validation process by randomly selecting 16 representative images from the test set. These images encompassed a broad spectrum of challenging scenarios, including umbrella occlusion, extreme lighting conditions (both overly bright and dim), substantial variations in target sizes, and intricate instances where target features closely resemble those of the background.

As depicted in Figure 11, our detection model triumphantly identified all targets within these intricate images. This performance can be attributed to the model’s effective feature fusion and representation capabilities, which empower it to dynamically modulate feature intensity and meticulously extract crucial information from multi-scale features. This adaptability enables the model to effectively discern targets from their surroundings, even in the most complex of scenarios.

Furthermore, the integration of multi-scale feature extraction and deep convolutional layers significantly bolstered the model’s representational power and computational efficiency. This enhancement allowed the model to seamlessly handle the challenges posed by varying target sizes and intricate environmental factors, ensuring consistent and accurate detections across the board.

In summary, the actual detection results presented in this section underscore the exceptional performance of our detection model in real-world applications. By leveraging advanced techniques such as enhanced feature fusion, multi-scale feature extraction, and deep convolutional processing, our model has demonstrated its ability to overcome even the most daunting of detection challenges, thereby validating its suitability for practical deployment in complex open channels.

3.4. Model Tracking Results and Analysis

3.4.1. MDRS Weight Ablation Experiment

To validate the scientific rationale behind the weight assignment in the Multi-Dimensional Re-Identification Scoring (MDRS) mechanism, we conducted a systematic ablation study on the validation set. This experiment aimed to quantitatively evaluate the contribution of each feature dimension to cross-camera target re-identification performance and determine the optimal weight combination through empirical analysis.

The proposed MDRS mechanism employs handcrafted low-level features and a fixed-weight fusion scheme, prioritizing computational efficiency and deployment simplicity for real-time, multi-camera industrial surveillance. While deep learning-based Re-ID embeddings offer superior invariance to appearance changes, their computational cost and need for extensive, labeled training data can be prohibitive for lightweight, scalable deployment on edge devices. The fixed weights, optimized empirically on our validation set, provide a stable and effective solution for the operational environment under study, where lighting variations are bounded (e.g., by structured artificial lighting at night). We acknowledge that in scenarios with extreme, unconstrained photometric variations, an adaptive weighting strategy or learned embeddings could further enhance robustness, which is a valuable direction for future work aimed at universal generalization.

A baseline configuration with uniform weights (1/6 ≈ 0.167 for each feature) was established to provide a neutral reference point, where no single feature dimension received preferential treatment. This baseline allows for clear measurement of performance improvements achieved through optimized weight allocation, As shown in

Table 6. MDRS Weight Ablation Study Results.

Configuration	Color	Texture	Shape	Motion	Size	Temporal	IDF1	Improvement over Baseline
Baseline	0.167	0.167	0.167	0.167	0.167	0.167	0.741	-
Color-only	1.0	0	0	0	0	0	0.782	+5.5%
No temporal constraint	0.42	0.21	0.16	0.16	0.05	0	0.801	+8.1%
Configuration B	0.35	0.25	0.15	0.15	0.05	0.05	0.815	+10.0%
Configuration C	0.45	0.15	0.15	0.15	0.05	0.05	0.821	+10.8%
Configuration D	0.40	0.20	0.10	0.20	0.05	0.05	0.812	+9.6%
Configuration E	0.40	0.20	0.20	0.10	0.05	0.05	0.808	+9.0%
Configuration A (Ours)	0.40	0.20	0.15	0.15	0.05	0.05	0.829	+11.9%

.

The experimental results provide several key insights into feature importance for cross-camera re-identification. Configuration A emerged as the optimal weighting scheme, achieving the highest IDF1 score of 0.829, which represents an 11.9% improvement over the uniform baseline.

The color-only configuration demonstrated that color features alone provide substantial discriminative power (+5.5% over baseline), confirming their status as the most stable and distinctive feature in port environments. However, the superior performance of Configuration A over the color-only approach highlights the importance of complementary features.

Comparative analysis of the configurations reveals the delicate balance required in weight assignment. Configuration B (emphasizing texture) and Configuration C (emphasizing color) both showed respectable improvements (+10.0% and +10.8%, respectively) but fell short of the optimal balance achieved in Configuration A. The performance degradation in Configuration D (reduced shape weight) and Configuration E (reduced motion weight) further underscores the importance of maintaining appropriate weights for these complementary features.

The critical role of motion consistency is evident from the 2.9% performance gap between Configuration A and Configuration E, demonstrating that motion features provide essential spatiotemporal context for cross-camera association. Similarly, texture and shape features serve as vital complements to color information, with Configuration A’s assignment of 0.20 and 0.15 weights proving optimal.

Size and temporal features function effectively as regularization terms with their low weights (0.05 each). The removal of temporal constraints resulted in a 2.8% performance drop compared to Configuration A, demonstrating their importance in filtering physically implausible associations across extended time intervals.

This systematic ablation study provides empirical justification for our weight selection and confirms that Configuration A establishes an optimal balance that emphasizes the most discriminative features while maintaining complementary support from secondary features and necessary regularization terms. The proposed weighting scheme effectively captures the relative importance of each feature dimension for cross-camera re-identification in complex port environments.

3.4.2. Tracking Results for Vehicles

This section evaluates the tracking performance of our proposed CSPC-BRS framework under both single-camera and cross-camera scenarios. The tracking requirements differ based on vehicle type and operational context: for smaller vehicles like electric bicycles and tricycles, the primary concern is monitoring safety compliance (e.g., helmet usage, unauthorized entry) within localized areas, making single-camera tracking sufficient. However, for trucks involved in cargo transportation, ensuring operational safety requires continuous monitoring across multiple zones, necessitating robust cross-camera identity consistency.

Figure 12 presents a comparative analysis of tracking efficacy in a low-light, backlit open environment, evaluating our CSPC-BRS framework against the baseline YOLOv8 + Bot-SORT. Under these challenging conditions, our CSPC-BRS framework successfully tracked all targets, including fast-moving bicycles and tricycles, maintaining a stable maximum ID count of 3 without any target loss. In stark contrast, the baseline model not only failed to correctly classify targets—misidentifying a bicycle as a person at frame 57—but also completely lost the trajectory of this target by frame 62. This misclassification and subsequent tracking failure underscore the baseline’s limitations in handling complex visual scenarios.

The performance disparity is further evidenced in Figure 13, where the baseline model consistently failed to detect the person target throughout the sequence. Our CSPC-BRS, however, reliably maintained all tracks with high confidence and without any identity switches, even as the tricycle underwent increased occlusion. These results collectively demonstrate the superior robustness of our method in adverse lighting conditions and its enhanced capability for handling rapid motion and partial occlusions.

The superior performance stems from the integrated CASAM attention mechanism, SPPELAN-DW aggregation network, and CACC convolutional modules, which collectively enhance feature extraction and fusion capabilities. Additionally, the BRS dynamically adjusts bounding box dimensions in response to velocity changes, significantly reducing target loss and ID switching.

To evaluate performance in cross-camera scenarios essential for monitoring cargo transportation safety, we designed an experiment tracking a truck across three surveillance cameras with significantly different viewpoints. Two cross-camera routes were established: Route 1 (camera1 to camera3) and Route 2 (camera2 to camera3).

As shown in Figure 14, CSPC-BRS maintained consistent ID assignment (ID = 1) for the truck throughout both routes, despite significant viewpoint variations and visual distortions. In Route 1, when the truck re-emerged after temporary occlusion by a gantry crane, the baseline model failed to re-identify the target, assigning new IDs (2–4). In contrast, CSPC-BRS successfully preserved the original identity, demonstrating the effectiveness of its multi-feature Re-ID scoring mechanism and BRS-based scale adaptation.

The baseline algorithm consistently assigned new IDs when the truck appeared in camera3 from both routes, indicating limited capability in handling appearance changes across viewpoints. CSPC-BRS not only maintained ID continuity but also sustained high-confidence bounding boxes throughout the tracking sequence, validating its robust performance in dynamic multi-camera environments.

This experiment confirms the effectiveness of our multi-dimensional Re-ID mechanism in complex cross-view scenarios and demonstrates the BRS module’s crucial role in compensating for scale variations, collectively enhancing tracking stability in practical logistics surveillance applications.

3.4.3. Evaluation of Employee Tracking in Open Channels

To quantitatively evaluate the tracking robustness of the proposed CSPC-BRS framework, we conducted a comparative study against the baseline YOLOv8+ Bot-SORT under diverse challenging scenarios in open channels, with a focus on identity (ID) preservation and occlusion handling.

The superior performance of our method is vividly demonstrated in Figure 15, which encapsulates tracking outcomes across two representative scenarios. In a scenario with frequent inter-person occlusion, the baseline model exhibited significant fragility. As individuals crossed paths, the baseline system suffered from multiple identity switches, causing its maximum ID count to increase by 33.3% (from 6 to 8), indicating a failure to maintain consistent trajectories. In stark contrast, our CSPC-BRS framework, empowered by the BRS, successfully maintained tracking continuity through the occlusion events, reducing the incidence of identity loss by 50% and demonstrating robust data association.

The advantage of our model becomes even more pronounced in highly complex environments characterized by strong illumination and persistent occlusions. In such demanding conditions, the baseline model struggled severely with both detection and tracking, failing to reliably identify safety helmets and suffering from rampant ID fragmentation. This led to a drastic 100% increase in the maximum ID count (from 12 to 24), rendering the tracking output unusable. Our proposed framework, however, proved highly resilient. It achieved stable detection despite the harsh lighting and, most importantly, successfully re-identified targets after prolonged occlusions (e.g., from frame 88 to 104), thereby limiting the maximum ID count to 13 and ensuring high-fidelity trajectory preservation.

In summary, the experimental results from multiple scenarios conclusively demonstrate that the integrated CSPC-BRS framework significantly enhances tracking stability and reliability. It substantially mitigates the critical challenge of identity switches in complex industrial settings, providing a more dependable solution for real-world safety monitoring applications.

3.5. Comprehensive Analysis of Actual Application Outcomes

To rigorously evaluate the practical efficacy of the CSPC-BRS framework, a controlled three-month comparative study was conducted using identical surveillance footage from four open-channel locations (A–D), with a focus on the number of false captures (FC) and the false capture rate (FCR). Both the baseline (YOLOv8 + Bot-SORT) and our proposed system processed the same video streams, ensuring a fair comparison under consistent operational conditions.

The CSPC-BRS framework was configured to automatically log and archive violation screenshots for subsequent analysis. Representative examples of these detections—including instances of missing safety helmets—are provided in Figure 16, visually affirming the model’s capability to identify genuine violations in complex port settings.

As summarized in

Table 7. Comparison of Results for Capturing Violation Images in Practical Applications.

Places	Violations	FC	FCR (%)	Violations	FC	FCR (%)	FCR Reduction Rate (%)
	Base Model (YOLOv8 + Bot-SORT)			Our Model (CSPC-BRS)
A	2122	67	3.16	1797	26	1.45	54.1
B	1808	87	4.81	1635	31	1.89	60.7
C	1460	78	5.35	1320	24	1.81	66.2
D	1246	45	3.61	1112	17	1.52	57.9
average							59.7

, CSPC-BRS reported fewer total violations across all locations. For instance, at Location A, violations decreased from 2122 (baseline) to 1797 (CSPC-BRS)—a reduction of 15.3%. Crucially, this decline is attributable not to an increase in missed detections, but to a significant drop in false positives and duplicate entries resulting from identity switches and tracking instability in the baseline model.

The baseline system’s propensity for identity fragmentation often registered a single violation event multiple times, artificially inflating its counts. In contrast, CSPC-BRS integrates CASAM for robust feature extraction, SPPELAN-DW for multi-scale stability, CACC for feature purification, and the BRS for motion-aware size adaptation. This synergy ensures consistent identity preservation and suppresses spurious or redundant detections.

To quantitatively address the concern of missed detections, we evaluated the false capture rate, which reflects the proportion of incorrect alerts. As shown in Table 7, across the four monitored locations (A–D), the proposed CSPC-BRS model achieved a significant reduction in the false capture rate compared to the baseline (YOLOv8 + Bot-SORT), with an average reduction of 59.7%. This confirms that the lower violation count stems from improved precision and tracking stability. Furthermore, manual audits of randomly sampled video segments throughout the deployment period did not reveal any noticeable increase in undetected helmet violations, corroborating that CSPC-BRS maintains high recall while significantly boosting precision.

Therefore, the observed reduction in violation counts reflects higher statistical accuracy and operational reliability. By minimizing false alarms and duplicate records, CSPC-BRS delivers a more credible and actionable safety monitoring system for port logistics management.

4. Limitations and Future Work

Despite the demonstrated effectiveness of the CSPC-BRS framework in the targeted port environment, several inherent limitations of the current work warrant discussion, which also define clear pathways for subsequent research. A primary limitation stems from the design of the Multi-Dimensional Re-Identification Scoring (MDRS) mechanism, which relies on a manually configured, static weighting scheme for feature fusion. While this configuration proved optimal on our validation set, its fixed nature may not dynamically adapt to more severe or unforeseen variations in imaging conditions, such as drastic changes in illumination spectrum between day and night or under intense weather effects, potentially impacting long-term robustness in fully unconstrained deployments. Furthermore, the experimental validation, while rigorous, is conducted within a domain-specific context derived from a single port’s operational data. The generalizability of the performance to geographically different logistics hubs, diverse seasonal scenarios, or other types of open industrial complexes remains a question that requires empirical verification. Lastly, the practical deployment scenario poses another consideration; the current performance metrics are established on a standard GPU platform, and the framework’s efficiency on resource-constrained edge devices, which are typical for widespread industrial IoT deployment, has not yet been evaluated, highlighting a gap between laboratory validation and field application.

To address these points, future research will focus on several targeted directions. The first direction involves evolving the MDRS into a more adaptive re-identification system. This could be achieved by investigating condition-aware weighting networks or integrating lightweight learned feature embeddings to replace or supplement the handcrafted features, thereby enhancing robustness to environmental dynamics. The second direction is to rigorously evaluate and improve cross-domain generalization. This will entail constructing a more diverse, multi-source benchmark dataset encompassing various ports and conditions, and exploring domain adaptation or generalization techniques to ensure the model’s reliability across heterogeneous environments. The third direction is oriented towards engineering for practical deployment, which includes porting the CSPC-BRS framework to mainstream edge computing platforms and applying model compression techniques such as pruning, quantization, and knowledge distillation. This effort aims to achieve an optimal balance between accuracy and inference speed under strict hardware constraints, facilitating real-world, on-device implementation.

5. Conclusions

This study proposed CSPC-BRS, a task-oriented framework for real-time multi-target detection and tracking in complex open-channel environments, with a particular focus on safety helmet monitoring in port logistics scenarios. By integrating the CSPC detector (composed of CASAM, SPPELAN-DW, and CACC) with the Bounding Box Reduction Strategy (BRS) and Multi-Dimensional Re-identification Scoring (MDRS), the proposed system effectively addresses critical operational challenges, including severe target scale variation, frequent occlusion, background interference, and identity instability during prolonged monitoring.

The introduced CSPC modules collaboratively enhance feature discrimination, scale stability, and cross-scale semantic consistency, leading to improved detection robustness under cluttered and low-quality visual conditions. At the tracking level, the BRS mechanism, together with motion-aware prediction and identity-aware scoring, significantly suppresses bounding box jitter and identity switches, resulting in more stable and reliable trajectory continuity.

Comprehensive evaluations, including ablation studies, quantitative comparisons with mainstream detectors, and a three-month real-world deployment, demonstrate the effectiveness of the proposed framework. CSPC-BRS achieves superior detection accuracy compared to the YOLOv8 baseline while maintaining real-time processing capability suitable for safety-critical monitoring applications. Notably, the system reduces the false capture rate by an average of 59.7% in practical deployment, indicating that the decline in recorded violations primarily stems from the elimination of spurious and duplicate alerts rather than missed detections. This improvement directly enhances the credibility and operational value of automated safety supervision systems.

It should be emphasized that the current evaluation is conducted under a domain-consistent setting reflecting the operational conditions of the investigated port environment. While the results demonstrate strong robustness within this context, large-scale multi-site and multi-season validation remains an important direction for further assessment of cross-domain transferability. Future work will also focus on deploying and benchmarking the framework on edge computing platforms (e.g., NVIDIA Jetson series) to validate its effectiveness under resource-constrained conditions, as well as exploring lightweight optimization strategies such as knowledge distillation and neural architecture search.

Overall, CSPC-BRS represents a robust and efficient solution for real-time safety monitoring in complex industrial environments. The experimental results indicate that the proposed framework is scalable and adaptable, supporting the development of intelligent safety management systems in port logistics and other related visual surveillance applications.