Impact of Water Velocity on Litopenaeus vannamei Behavior Using ByteTrack-Based Multi-Object Tracking

Abstract

In factory-controlled recirculating aquaculture systems, precise regulation of water velocity is crucial for optimizing shrimp feeding behavior and improving aquaculture efficiency. However, quantitative analysis of the impact of water velocity on shrimp behavior remains challenging. This study developed an innovative multi-objective behavioral analysis framework integrating detection, tracking, and behavioral interpretation. Specifically, the YOLOv8 model was employed for precise shrimp detection, ByteTrack with a dual-threshold matching strategy ensured continuous individual trajectory tracking in complex water environments, and Kalman filtering corrected coordinate offsets caused by water refraction. Under typical recirculating aquaculture system conditions, three water circulation rates (2.0, 5.0, and 10.0 cycles/day) were established to simulate varying flow velocities. High-frequency imaging (30 fps) was used to simultaneously record and analyze the movement trajectories of Litopenaeus vannamei during feeding and non-feeding periods, from which two-dimensional behavioral parameters—velocity and turning angle—were extracted. Key experimental results indicated that water circulation rates significantly affected shrimp movement velocity but had no significant effect on turning angle. Importantly, under only the moderate circulation rate (5.0 cycles/day), the average movement velocity during feeding was significantly lower than during non-feeding periods (p < 0.05). This finding reveals that moderate water velocity constitutes a critical hydrodynamic window for eliciting specific feeding behavior in shrimp. These results provide core parameters for an intelligent Litopenaeus vannamei feeding intensity assessment model based on spatiotemporal graph convolutional networks and offer theoretically valuable and practically applicable guidance for optimizing hydrodynamics and formulating precision feeding strategies in recirculating aquaculture systems.

1. Introduction

Aquaculture has become a cornerstone of global food security, representing one of the fastest-growing sectors in global food production. It plays a critical role in supplying aquatic products, which are a primary source of protein for humans [1]. Driven by technological innovation and market demand, China’s aquaculture industry has expanded significantly. In 2024, China’s total aquaculture production exceeded 73.66 million metric tons, maintaining its position as the global leader for consecutive years [2]. The Pacific white shrimp (Litopenaeus vannamei) has demonstrated outstanding performance in aquaculture due to its unique biological characteristics [3]. Since its introduction to China, this species has shown significant advantages, including rapid growth, strong environmental adaptability, and high feed conversion efficiency [4,5]. Recirculating aquaculture systems, characterized by controllable water quality, high-density intensive culture, and environmental friendliness, are progressively replacing traditional open-pollinated farming models, becoming a core technology for the green transformation of the shrimp industry [6]. Traditional aquaculture relies on large water body exchanges, leading to issues such as water resource waste, frequent disease outbreaks, and environmental pollution [7]. In contrast, recirculating aquaculture systems utilize modular designs like biological filtration and protein separation to precisely maintain stable water quality parameters (e.g., ammonia nitrogen, nitrite), significantly reducing stress-induced feeding inhibition in shrimp [8]. Furthermore, its closed-loop nature allows for over 90% reduction in wastewater discharge [9]. Under recirculating aquaculture system conditions, shrimp feeding behavior is regulated by a synergistic interplay of factors, including water quality, light cycle, feed properties, and water velocity [10,11,12,13]. Among these, water velocity has become a specific focus of research due to its dual mechanisms of action: (1) It directly regulates prey diffusion efficiency and shrimp feeding posture [14], modifies energy allocation for locomotion, and influences social interactions. However, the ecological mechanisms behind such adaptations remain less explored than in fish studies, partly due to methodological limits. Bardera et al. (2019) shed light on this by proposing an energy conservation paradigm, showing that penaeid shrimp in low-shear environments (<0.1 m/s) minimize locomotion costs to optimize foraging—prioritizing digestion over mobility, unlike fish metabolic compensation [14]. (2) It indirectly affects net feeding efficiency by influencing oxygen distribution and the transport of metabolic waste [15]. Quantitative analysis of its independent effects and interaction mechanisms is essential for providing theoretical support for precision feeding and energy optimization in recirculating aquaculture systems.

Water velocity is a critical environmental factor regulating shrimp feeding behavior, with direct effects evident in feeding frequency, efficiency, and physiological responses [16]. Research indicates that moderate water velocity can enhance shrimp activity space and foraging capability, thereby significantly increasing feeding frequency and biomass [17]. Concurrently, it optimizes the feed conversion ratio and reduces feed waste [18]. Furthermore, water velocity interacts significantly with other environmental factors, such as dissolved oxygen and temperature. For instance, under low dissolved oxygen conditions, moderate flow can promote oxygen diffusion throughout the water column, thereby improving respiration and potentially boosting feeding [19]. From a physiological perspective, extreme velocities can trigger stress responses, leading to energy allocation imbalances and immune suppression [20]. Conversely, appropriate flow rates help maintain physiological homeostasis and enhance nutrient absorption efficiency by increasing digestive enzyme activity [18]. These studies collectively reveal the complex mechanisms through which water velocity influences shrimp feeding via both behavioral regulation and physiological adaptation.

In the field of intelligent shrimp behavior monitoring, computer vision technology has become the core methodology, enabling the transformation of image or video features into quantitative behavioral data [21]. Behavior recognition based on computer vision typically integrates key algorithms such as image preprocessing, segmentation, and feature extraction [22]. Among these, methods employing region segmentation/edge detection, threshold processing, and background subtraction for behavior area delineation, individual identification, and precise localization have become the mainstream paradigm in current behavior tracking and detection research. The introduction of deep learning technology has significantly advanced this field. Deep learning-based object detection algorithms (e.g., YOLO series) have effectively addressed the limitations of traditional methods, achieving automatic shrimp identification in complex aquaculture environments and demonstrating high accuracy and industrial application potential [23]. However, although various behavior tracking software packages (such as idtracker.ai, Tracktor) have been developed, these tools are primarily designed for multi-object tracking in short videos. They often lack effective analytical methods for long-duration behavioral video data, making them inadequate for the precise analysis of individual shrimp trajectories over extended periods, which is crucial for factory-scale aquaculture scenarios. Therefore, developing robust multi-object tracking models capable of overcoming challenges such as small target size, high density, and individual identity switching (ID switching), while being suitable for long-duration video analysis, is key to advancing intelligent shrimp behavior research and supporting precision aquaculture management.

This study constructed a computer vision analysis system based on the ByteTrack multi-object tracking algorithm to simultaneously extract movement parameters (speed and turning angle) of Litopenaeus vannamei through spatiotemporal feature fusion. By setting up experimental groups with different water recirculation rates (gradients: 2, 5, and 10 cycles/day), the system analyzed the impact mechanism of flow velocity variations on the schooling flow-following behavior of shrimp. Recent studies have validated the efficacy of ByteTrack-based frameworks combined with spatiotemporal graph convolutional networks (ST-GCNs) for robust behavior quantification in aquaculture, particularly in resolving challenges such as target occlusion and identity switching under high-density conditions [24,25]. Based on the experimental results, an optimal water recirculation rate was selected, which was then employed for the intelligent assessment of shrimp feeding intensity using spatiotemporal graph convolutional neural networks (ST-GCNs). This research provides a quantified basis with both theoretical value and engineering applicability for optimizing hydrodynamic parameters in recirculating aquaculture systems, thereby supporting dynamic matching of feeding strategies and control of energy consumption costs.

2. Materials and Methods

2.1. Experimental Materials and Daily Management

The culture experiment was conducted at the Langya Experimental Base of the Yellow Sea Fisheries Research Institute, Chinese Academy of Fishery Sciences (Qingdao West Coastal New Area). Five hundred shrimp (body length: 15 ± 3 cm) were selected for the experiment, with a stocking density of approximately 5 kg/m³. The aquaculture water was natural seawater treated with sedimentation and filtration. During the experiment, the water temperature was maintained at 25 ± 1 °C, salinity at 30 ± 1‰, and dissolved oxygen (DO) at >5.0 mg/L through continuous aeration using aerators. Additionally, a temperature control system was employed to stabilize the water temperature at (24.5 ± 1) °C. Daily water exchange accounted for 10% of the total volume, concurrently removing settled feed residues and excreta from the bottom of the culture tanks. The filter media in the filtration system were also regularly replaced to ensure stable water quality.

2.2. Dataset Collection and Annotation

To investigate the influence of recirculating water flow speed on shrimp movement behavior, this study systematically collected experimental video data encompassing various behavioral states. For quantitative behavioral analysis, the original videos were first processed to extract keyframes using a sampling strategy of extracting 1 frame every 300 frames. This approach ensured sufficient temporal gaps between images while preserving complete behavioral features as much as possible. The downsampled image samples included two typical behavioral states: feeding and non-feeding. Non-feeding state samples were collected before feed delivery, whereas feeding state samples recorded the feeding process of the shrimp after feed delivery. This bimodal sample design effectively captured the differential impacts of recirculating water flow speed on movement characteristics across different behavioral modes, providing comprehensive data support for subsequent temporal behavioral analysis.

Object detection, a core task in computer vision, aims to achieve precise identification and localization of objects within images. Constructing a high-quality object detection dataset requires meeting two key conditions: sufficient image sample quantity, and accurate annotation information—including the spatial location of each object marked by bounding boxes and its corresponding category. The shrimp detection dataset constructed in this study adopts the PASCAL VOC standard annotation format. This format stores annotation information collaboratively in XML files paired with corresponding image files. The XML files structurally record image resolution, object categories, and the geometric coordinates (top-left and bottom-right corner values) of each annotation box. The open-source tool Labellmg was selected for interactive annotation tasks. This tool supports visual selection of target regions and the generation of standard VOC format annotation files. Additionally, it includes an integrated format conversion module that can automatically convert XML annotations into the normalized data format required for YOLO models (i.e., class index, normalized center x-coordinate, normalized center y-coordinate, box width, and box height).

2.3. Experimental System Configuration

To investigate the effects of recirculating water velocity on shrimp behavior in industrial aquaculture settings, this study designed and constructed an experimental apparatus based on an operational recirculating aquaculture system. The laboratory-built system (Figure 1) comprises two primary subsystems: the recirculating aquaculture system and the monitoring system, with the following components:

Culture Tank: Primary site for shrimp cultivation and behavioral observation. Constructed as a custom acrylic circular tank (Ø 4.5 m; water depth 0.6 m; total volume 10 m³). The experimental tank employs a tangential water inlet combined with a central bottom-drain outlet configuration, which generates a rotational flow field in the circular tank. The tangential inlet induces high-momentum circumferential currents along the tank walls, while the central drain creates a low-pressure zone that promotes radial inflow toward the tank center. This synergy results in a characteristic velocity dichotomy: high shear zones near the periphery versus quiescent regions at the core.

Drum Filter (Microscreen): Core physical filtration unit removing suspended solids (feed residue, feces, algae) with 10–200 μm filtration precision.

Biofilter: Central bio-remediation unit employing microbial nitrification to convert toxic ammonia (NH₃/NH₄⁺) to nitrate (NO₃⁻). This engineered microbial ecosystem is essential for recirculating aquaculture system water treatment.

Thermostatic System: Critical module maintaining water temperature at 25 ± 0.5 °C.

Aeration System: To ensure the independent impact of flow velocity in the experiment, this study strictly controlled the dissolved oxygen (DO) conditions of the aquaculture water through an aeration system. During the experiment, the DO content was stably maintained at 5.0–7.0 mg/L, which was significantly higher than the feeding inhibition threshold (3.0 mg/L) and stress threshold (2.0 mg/L) of Litopenaeus vannamei. The shrimp did not show abnormal behaviors related to hypoxia (such as floating heads and disordered swimming). Therefore, the behavioral differences between different experimental groups can be clearly attributed to the changes in flow velocity rather than the interference of DO fluctuations.

Ultra-Wide-Angle Camera (Hikvision 3347WDP2V2-L, Hikvision, Hangzhou, China): Imaging device capturing shrimp behavior at 30 fps (1920 × 1080 resolution). Videos stored in MP4 format.

Monitoring Host (Hikvision, Hangzhou, China): Central unit for data acquisition, real-time monitoring, and storage.

Rotary-Current Meter (OKA LS-300A, Qingdao Jingcheng Instrument Co., Ltd., Qingdao, China): Portable velocimeter measuring flow velocity via turbine rotation. Enables precision mapping of spatial velocity gradients (center vs. peripheral zones) for hydraulic/environmental applications.

Figure 1. Experimental setup (adapted from Huang et al., 2024 [26]).

Figure 1. Experimental setup (adapted from Huang et al., 2024 [26]).

2.4. ByteTrack Algorithm Overview

2.4.1. Algorithm Execution Flow

The core objective of object detection and tracking algorithms is to construct trajectories for detected objects and form complete trajectories by matching detection boxes frame by frame. The specific flow of the ByteTrack algorithm is illustrated in Figure 2:

• The YOLOv8 model is utilized to perform object detection on the fish school in the culture tank. Based on confidence scores, detection boxes are categorized into high-confidence and low-confidence classes.
• Initial trajectory matching is performed on high-confidence detections. This involves calculating their Intersection over Union (IoU) with existing trajectories and applying the Hungarian algorithm for matching. Successfully matched trajectories are updated via Kalman filtering and added to the current frame’s trajectory set. Unmatched trajectories and detection boxes are respectively stored in the set of unmatched trajectories for further processing.
• A second round of IoU matching is conducted between low-confidence detections and the previously unmatched trajectories. Successfully matched trajectories are reactivated and updated. Unmatched low-confidence detections are considered background noise and are deleted. During trajectory management, new trajectories are created and added to the current frame’s trajectory set for detection boxes with confidence scores above the tracking threshold. Additionally, trajectories that remain unmatched for 30 consecutive frames are deleted. Finally, the current frame’s trajectory set is passed to the next frame for Kalman filtering to predict new positions. Through these steps, the ByteTrack algorithm achieves an efficient combination of detection and tracking, significantly improving the accuracy and stability of trajectories (the flow of the ByteTrack algorithm is illustrated in Figure 2).

2.4.2. Kalman Filtering

Kalman filtering is an optimal estimation algorithm based on linear system state equations. It estimates the state by combining system input and output data. Under conditions where observation data is subject to noise interference, it can extract hidden state information and predict future states based on historical data. This algorithm is characterized by low memory usage and fast computation speed, making it suitable for real-time processing and embedded systems. It is widely applied in fields such as target tracking, navigation, control, signal processing, econometrics, and robotics.

The core idea of Kalman filtering is to combine the mathematical model of a system with its observation data. The mathematical model describes the dynamic evolution of the system’s state, while the observation data provides state information corrupted by noise. The algorithm updates the state estimate and its uncertainty recursively by utilizing both the model and the observations. Under the assumption that errors follow a normal distribution, Kalman filtering provides the optimal estimate with the minimum mean square error. Its basic procedure involves two steps: first, predicting the current state based on the system’s dynamic model, and second, correcting the prediction result by combining the observation model and the actual observation data, thereby obtaining the optimal estimate. This process is recursive, where each state estimate depends on the previous estimate and the current observation data.

The core of the Kalman filter is composed of two sets of equations: the prediction equation and the update equation. The prediction equation describes the dynamic behavior of the system state, while the update equation is used to correct the state estimate. These two sets of equations together form the fundamental framework of the Kalman filter. The general form of the Kalman filter equations is as follows:

Prediction equation:

$${\widehat{x}}_k^{-}=F_k{\widehat{x}}_{k-1}+B_k{u}_k$$

(1)

$$P_k^{-}=F_k{P}_{k-1}{F}_k^T+Q_k$$

(2)

Update equation:

$$K_k=P_k^{-}{H}_k^T{\left(H_k{P}_k^{-}{H}_k^T+R_k\right)}^{-1}$$

(3)

$${\widehat{x}}_k={\widehat{x}}_k^{-}+K_k\left(z_k-H_k{\widehat{x}}_k^{-}\right)$$

(4)

$$P_k=\left(I-K_k{H}_k\right)P_k^{-}$$

(5)

Matching criterion equation for high- and low-confidence detection boxes:

$$IoU\left(a,b\right)={\displaystyle \frac{\left|a\cup b\right|}{\left|a\cap b\right|}}$$

(6)

Motion constraint equation:

$$v_{min}\le ‖\overrightarrow{v}‖\le v_{max}$$

(7)

In this context, (${\widehat{x}}_k$) represents the state estimate at time step (_k), (P_k) is the state estimate error covariance matrix at time step (_k), (F_k) is the state transition matrix, (B_k) is the control matrix, (u_k) is the control input, (Q_k) is the process noise covariance matrix, (H_k) is the observation matrix, (z_k) is the observation data, (R_k) is the observation noise covariance matrix, and (K_k) is the Kalman gain.

Equation (1) (State Prediction): Describes the temporal evolution of shrimp movement states. Based on the previous state, including position and velocity, it predicts the next state via the state transition matrix, reflecting the inertial continuity of their swimming while incorporating control inputs like limb movement thrust.

Equation (2) (Uncertainty Propagation): Quantifies prediction uncertainty. It derives propagated uncertainty from the prior state’s error covariance and state transition matrix, plus environmental disturbances like water flow fluctuations, reflecting the impact of velocity gradients in RAS on prediction accuracy.

Equation (3) (Kalman Gain): Balances prediction uncertainty (from water flow and other disturbances) and observation noise (e.g., water refraction errors), determining the weight of observation data in state correction.

Equation (4) (State Update): Corrects the predicted state using the deviation (innovation) between observations and predictions, offsetting errors caused by water flow disturbances or random shrimp movement to make estimates closer to actual positions.

Equation (5) (Covariance Update): Updates the confidence of state estimates. The more reliable the observations (e.g., clear images), the smaller the error covariance, reflecting reduced uncertainty after correction.

Equation (6) (IoU Matching): Quantifies spatial overlap between trajectory boxes (a) and low-confidence detection boxes (b) by calculating their Intersection over Union. A match is confirmed when IoU ≥ 0.3, enabling trajectory recovery of occluded targets and resolving ID switches in high-density aggregations.

Equation (7) (Motion Constraint): Defines the physical range [1.2, 35.6 pixel/s] for shrimp’s instantaneous velocity ($‖\overrightarrow{v}‖$), filtering outliers from water refraction or imaging noise. This provides stable input for Kalman filtering, adapting to the actual movement capability of shrimp.

2.4.3. Hungarian Algorithm

The Hungarian algorithm, a classic method in the field of combinatorial optimization, fundamentally revolves around constructing a mathematical framework for finding optimal matchings in bipartite graphs. Its objective is to find an optimal matching in a bipartite graph that minimizes or maximizes the sum of the weights of the matched edges. This algorithm is not only applicable to classical assignment problems but can also be extended to complex scenarios such as multi-objective assignment problems and quadratic assignment problems. Its theoretical value lies in ensuring global optimality through an augmenting path search strategy, while its practical advantage lies in its robustness, making it capable of meeting the computational requirements of real-time control systems and edge computing devices.

The Hungarian algorithm can be represented in matrix or graph form. Given a non-negative matrix (C) of size (n × n), where element (C_ij) represents the cost of assigning worker (i) to task (j), the algorithm aims to find an assignment scheme such that each worker is assigned exactly one task, each task is performed by exactly one worker, and the total cost is minimized. If the goal is to maximize the total cost, the cost matrix can be transformed by (C_ij = \text {max cost} − C_ij). The specific steps of the algorithm are as follows: First, subtract the minimum element from each row of matrix (C) to obtain a new matrix (C′), ensuring that each row has at least one zero element. Next, subtract the minimum element from each column of matrix (C′) to obtain a new matrix (C″), ensuring that each column has at least one zero element. Then, cover all zero elements in matrix (C″) using the minimum number of horizontal and vertical lines, and record the number of covering lines (k). If (k = n), the optimal assignment scheme has been found, and the algorithm terminates. Otherwise, find the minimum value (m) among the uncovered elements. Subtract (m) from each uncovered element and add (m) to each element covered twice to obtain a new matrix (C‴). Finally, repeat the above steps until the optimal assignment scheme is found.

In this study, an innovative three-stage analysis framework of Detection-Tracking-Behavioral Analysis is constructed. It enhances the robustness of small object detection through the YOLOv8-SPD convolutional module, ensures trajectory continuity in occluded scenarios based on ByteTrack’s dual-threshold matching strategy, and analyzes the impact of locomotor activity cycles on the feeding and non-feeding states of Litopenaeus vannamei by examining kinematic parameters.

2.4.4. Technical Features of the Integrated MOA Framework

A key technical choice in the MOA framework is the adoption of YOLOv8 for shrimp detection, with three critical advantages tailored to aquaculture scenarios:

First, it excels at detecting small and occluded targets. In RAS environments, shrimp often appear as small targets in high-resolution images and are frequently occluded by bubbles, feces, or conspecifics. YOLOv8′s optimized CSPDarknet backbone and PAFPN neck structure enhance multi-scale feature fusion, enabling it to preserve fine-grained details of partially obscured individuals—directly addressing the challenge of detecting such targets in turbid water. Second, it balances speed and precision for real-time tracking. Unlike methods relying on region proposal networks (which lead to slower inference), YOLOv8 adopts an anchor-free detection mechanism and end-to-end training, allowing it to process video at rates consistent with typical monitoring systems while maintaining high precision. This real-time performance is critical for subsequent tracking steps (e.g., ByteTrack), as delayed detection would disrupt trajectory continuity and degrade ID consistency. Third, it exhibits strong robustness to aquatic environmental noise. Water refraction, dynamic light conditions, and suspended particles in RAS introduce significant noise into images. YOLOv8′s improved loss function (CIoU loss) and data augmentation strategies (including underwater-specific augmentations like fog and motion blur in training) enhance its tolerance to such noise, ensuring stable detection accuracy across varying water clarity and reliable target initialization for tracking even in suboptimal imaging conditions.

The Multi-Objective Analysis (MOA) framework, built on the ByteTrack algorithm, incorporates three key technical adaptations for aquaculture scenarios: Enhanced detection via YOLOv8 ensures robust target initialization, particularly for small shrimp targets affected by water refraction or partially occluded by bubbles/feces; ByteTrack’s dual-threshold matching (detailed in Section 2.4.1) maintains trajectory continuity, with high-confidence detections (score > 0.8) matched to Kalman-predicted trajectories and low-confidence ones (0.5–0.8) undergoing secondary IoU matching to reduce ID switches during shrimp aggregation; and refraction-aware Kalman correction (based on 2.4.2) calibrates coordinate offsets from the air–water interface (n₁ = 1.000, n₂ = 1.333) for accurate extraction of kinematic parameters (velocity, turning angle). This integrated design enables precise, continuous tracking of Litopenaeus vannamei in complex RAS environments.

3. Results

3.1. Model Performance Evaluation Results

The object detection results using the YOLOv8 model show that it significantly outperforms the YOLOv5 and Faster R-CNN models in terms of Precision, Recall, F1-score, and mAP50, achieving scores of 90.0%, 87.6%, 88.3%, and 92.8%, respectively (

Table 1. Evaluation metric results for three target detection models.

Model	Precision (%)	Recall (%)	mAP50 (%)	F1-Score (%)
YOLOv8	90.0	87.6	92.8	88.3
Faster R-CNN	67.6	64.3	78.3	55.9
YOLOv5	84.8	75.6	83.1	79.9

). This indicates that the YOLOv8 model possesses higher classification accuracy and positive sample coverage.

3.2. Spatial Distribution Characteristics of Water Flow Velocity in Ponds Under Different Water Circulation Rate Gradients

During the experiment, water flow velocity was measured simultaneously in the pool center (radius ≤ 0.2 R) and the edge area (0.8 R ≤ radius ≤ R) using flowmeters, based on three control gradients of water circulation rate: low (2.0 times/day), medium (5.0 times/day), and high (10.0 times/day). The results showed the following: (1) For spatial difference, under all three water circulation rate conditions, the average flow velocity in the edge area was significantly higher than that in the center area, and (2) for the gradient response pattern, as the water circulation rate increased, the velocity difference between the center and the edge expanded from 0.0073 m/s to 0.0339 m/s (the flow velocities at the center and edge of the pond are presented in

Table 2. Center flow rates of the three recirculation volume gradient pools.

Test Point	Low Recirculation (m/s)	Medium Recirculation (m/s)	High Recirculation (m/s)
1	0.018	0.023	0.037
2	0.021	0.026	0.034
3	0.020	0.024	0.032

and

Table 3. Flow velocities at the outer edge of the three recirculation volume gradient pools.

Test Point	Low Recirculation (m/s)	Medium Recirculation (m/s)	High Recirculation (m/s)
1	0.025	0.045	0.078
2	0.030	0.044	0.072
3	0.025	0.043	0.070
4	0.029	0.041	0.057
5	0.028	0.039	0.056
6	0.025	0.039	0.076

, respectively).

3.3. Response Characteristics of Shrimp Feeding Behavior to Water Circulation Rate

This study aims to analyze the response of Litopenaeus vannamei to pond water circulation rates. It employs a dual-modal analysis framework (feeding state vs. non-feeding state) to systematically quantify differences in movement characteristics under various water circulation rates (2.0, 5.0, 10.0 cycles/day). Using a YOLOv8-ByteTrack multi-object tracking system, individual movement trajectories were accurately analyzed in actual industrial recirculating aquaculture systems (the trajectory visualization is shown in Figure 3). In Figure 3, the multi-object tracking visualization presents the movement trajectories of Litopenaeus vannamei under different water circulation rates (2.0, 5.0, and 10.0 cycles/day) during feeding and non-feeding states. The colored objects represent individual shrimp, with each unique color assigned to a specific shrimp to distinguish its trajectory from others throughout the tracking period. The numbers adjacent to the colored objects are individual identification codes, which uniquely mark each shrimp to ensure continuous tracking—even when temporary occlusion or position shifts occur, these codes allow for the consistent mapping of an individual’s complete movement path over time. The trajectories in different colors intuitively reflect the differences in movement ranges and paths among individuals within the population. By combining the identification codes, changes in individual behavior between feeding and non-feeding states can be observed, providing visual evidence for the subsequent finding that a significant difference in movement velocity between feeding and non-feeding states only occurs under the 5.0 cycles/day circulation rate. Statistical analysis was performed using one-way Analysis of Variance (ANOVA) to examine the effects of water circulation rate gradients (2.0, 5.0, and 10.0 cycles/day) on movement parameters (velocity and turning angle). Significant differences identified via ANOVA are annotated in corresponding figures using superscript letters (a, b, ab), where identical letters indicate no significant difference (p > 0.05) and distinct letters denote statistical significance (p < 0.05). The experimental results indicate that under a water circulation rate of 2 cycles/day, the average velocity of shrimp during non-feeding states was slightly higher (16.02 pixel/s) than during feeding states (15.58 pixel/s), but this difference was not statistically significant at the 0.05 level. Under a water circulation rate of 5 times/day, the average velocity during non-feeding states (23.57 pixel/s) was significantly higher than during feeding states (13.58 pixel/s), showing a statistically significant difference at the 0.05 level. Under a water circulation rate of 10 times/day, the average velocity during non-feeding states (13.99 pixel/s) was slightly higher than during feeding states (12.34 pixel/s), and this difference was also not statistically significant at the 0.05 level (as shown in

Table 4. Changes in kinematic eigenvalues at different cycle volumes.

Recirculation Volume		Average Speed (pixel/s)	Standard Deviation	Maximum Speed (pixel/s)	Maximum Speed (pixel/s)
2	Feeding	15.58	4.77	24.43	10.15
	Non-feeding	16.02	5.84	33.27	10.88
5	Feeding	13.58	3.17	22.36	10.17
	Non-feeding	23.57	2.87	27.56	17.38
10	Feeding	12.34	2.56	17.73	9.17
	Non-feeding	13.99	5.61	30.55	7.91

). These velocity patterns are visually reflected in Figure 4 and Figure 5, which present the movement parameters (velocity and turning angle) of Litopenaeus vannamei under feeding and non-feeding states across the three water circulation rates. In the velocity subplot in Figure 4, at 5.0 cycles/day, feeding velocity is marked with a distinct letter compared to non-feeding velocity, indicating a significant difference (p < 0.05), while at 2.0 and 10.0 cycles/day, the same letters denote no statistical difference (p > 0.05). Figure 5 further details these state-specific variations, with the velocity subplot highlighting the marked difference at 5.0 cycles/day through distinct letters for non-feeding and feeding states, and minimal variations at other rates with the same letters. Water circulation rate gradients significantly affected shrimp movement velocity, whereas turning angle parameters showed no statistically significant difference at the 0.05 level (as shown in Figure 4). Furthermore, movement velocity differed significantly between feeding and non-feeding states at the 0.05 level, while turning angle parameters did not show a statistically significant difference between these states at the 0.05 level (as shown in Figure 5). Specifically, in both Figure 4’s turning angle subplot and Figure 5’s turning angle subplot, all groups are labeled with the same letters, confirming that turning angles remained consistent across all rates and states, with no statistical differences (p > 0.05). Together, these figures visually reinforce the key findings that a significant reduction in feeding velocity only occurred at 5.0 cycles/day, while turning angles were unaffected by either water circulation rate or feeding state.

4. Discussion

This study integrates advanced computer vision technology with hydrodynamic environmental control to systematically analyze the regulatory mechanisms of water circulation rate (representing flow velocity) on the behavioral characteristics of Litopenaeus vannamei within industrial recirculating aquaculture systems. The discussion is structured into three key aspects: model performance, flow field characteristics, and behavioral responses.

4.1. Analysis of Model Performance Evaluation Results

The model performance evaluation revealed a significant advantage for the YOLOv8 model across key metrics: Precision (90.0%), Recall (87.6%), F1-score (88.3%), and mAP50 (92.8%). This represents an improvement of approximately 12–15% over YOLOv5 and 18–22% over Faster R-CNN, confirming the effectiveness of its architectural enhancements. This performance gain aligns with current trends in underwater object detection, where the YOLO series algorithms have become the mainstream architecture due to their balanced advantages in accuracy and speed [27]. Compared to YOLOv3, subsequent iterations (including the YOLOv8 model employed in this study) have consistently improved performance through architectural refinements, such as the SPD-Conv module adopted here for enhanced small object feature extraction and the Task-Aligned Assigner for optimized dense object detection [28]. Notably, the model’s Recall rate of 87.6% is crucial for the subsequent ByteTrack multi-object tracking, ensuring trajectory continuity in complex recirculating aquaculture system environments characterized by dense shrimp aggregations and transient occlusions [29]. While the YOLOv3-SPP model achieved an impressive mAP50 of 97.03% [28], the YOLOv8 model in this study attained comparable high accuracy under the more challenging conditions of actual aquaculture scenarios (e.g., water refraction, dynamic backgrounds, and lighting variations), validating its robustness and applicability in this research context. Future work could explore integrating channel-spatial attention mechanisms [30] or Spatial Pyramid Pooling (SPP) modules [28] to further improve the detection rates of small objects or low-contrast shrimp.

4.2. Analysis of the Spatial Distribution Characteristics of Flow Velocity in the Culture Tank

Flow velocity monitoring results indicated that the velocity difference between the edge and central regions of the culture tank significantly increased with increasing water circulation rate. Oca and Masalo’s angular momentum model (β∝e^kr) explains radial velocity gradients in circular tanks, which aligns with our observed edge–center velocity differences (Table 2 and Table 3). However, their focus on flow dynamics in empty or fish tanks overlooks biological modulation. Our study revealed that Litopenaeus vannamei actively reinforced low central velocities by reducing movement (Δv = 10 pixel/s, p < 0.05) within a species-specific window (0.03–0.06 m/s), introducing a bio-fluid coupling mechanism unaccounted for in their model [31]. This phenomenon is closely linked to the flow field characteristics inherent to recirculating aquaculture systems. From a fluid dynamics perspective, the tangential inflow design induces high-speed rotational flow along the tank walls [32], while the central region forms a low-velocity zone due to the balance between centrifugal force and pressure gradients [33]. The tangential inlet and central bottom-drain design of the tank creates a stable rotational flow field, which directly influences the shrimp’s swimming path selection and feeding activity frequency, revealing the intrinsic link between water flow patterns in aquaculture facilities and aquatic organism behavior. This velocity gradient distribution corresponds to the “tea cup effect” described in the literature, where rotating water flow generates an inward radial current on the tank bottom, driving particulate matter towards the central drain; however, the central region may develop “dead zones” due to insufficient velocity [34]. As the water exchange rate increased from 2.0 to 10.0 cycles/day, the velocity increase in the peripheral region was markedly greater than in the central region. This suggests that energy input primarily enhances the peripheral tangential flow, while modifications to the central flow field face physical constraints, potentially related to the conservation of angular momentum in rotating fluids [35]. The observed velocity gradient pattern (high velocity at the periphery, low velocity at the center) and its response to the water circulation rate (gradient widening with flow rate) are consistent with flow field results obtained from computational fluid dynamics (CFD) simulations and Particle Image Velocimetry (PIV) techniques for circular/octagonal recirculating aquaculture system tanks [36]. This spatial heterogeneity in flow velocity provides a crucial environmental context for understanding shrimp distribution and behavioral differences (e.g., feeding position preferences) across tank regions and highlights the need for special attention to water circulation efficiency in the central zone when optimizing recirculating aquaculture system design.

4.3. Analysis of the Response Characteristics of Shrimp Feeding Behavior to Water Circulation Rate

Shrimp behavioral observations revealed that under a water circulation rate of 2.0, 5.0, and 10.0 cycles/day, a significant reduction in average velocity during feeding states (13.58 pixel/s) compared to non-feeding states (23.57 pixel/s) occurred only at 5.0 cycles/day (42.4% reduction, p < 0.05), demonstrating distinct behavioral differences. This phenomenon stems from the non-linear coupling mechanism between hydrodynamic stress and feeding behavior. At the low circulation rate (2.0 cycles/day), the average edge flow velocity was only 0.028 m/s (Table 3), significantly below the threshold for solid waste removal (15–30 cm/s) [37]. This weak flow was insufficient to trigger energy conservation mechanisms in shrimp [38], resulting in an inability to kinematically distinguish feeding behavior from random swimming (velocity difference of only 0.44 pixel/s). While this aligns with Gorle et al.’s [38] observation of expanded dissolved oxygen gradients under low flow, the present study is the first to quantify its negative impact from the perspective of behavioral distinctiveness. Gorle et al. showed that salmon exhibited biomass-induced velocity attenuation (V_β = 0.75× V₀) and used metabolic compensation to maintain optimal swimming speeds (1–1.5 BL/s). In contrast, our shrimp prioritized feeding over mobility at ultra-low velocities (0.045 m/s)—a threshold 4–8× lower than that of salmon. Our 30 fps YOLOv8-ByteTrack framework, capturing metrics like turning stability, addresses their call for “higher-resolution behavioral mapping,” highlighting crustacean–fish differences in flow adaptation. Conversely, at the high circulation rate (10.0 cycles/day), the edge velocity of 0.070 m/s exceeded the threshold for postural control, inducing compensatory movement effects and confirming the phenomenon of “imbalanced energy allocation” [20], thereby keeping feeding period velocity (12.34 pixel/s) close to non-feeding period velocity (13.99 pixel/s). This finding exhibits parallels with observations in Portunus trituberculatus [39], where trajectory complexity significantly increased at 2.5 cm/s flow velocity, exhibiting parallel ethological patterns. The key insight lies in the 5.0 cycles/day rate aligning precisely with a “hydrodynamic window period”: This flow regime, through moderate hydrodynamic stress, induces a “feeding priority” strategy. It avoids the insufficient feed dispersion encountered at low flow velocities while preventing neuromuscular control overload at high velocities, thereby maximizing the difference in locomotor characteristics between feeding and non-feeding states (velocity difference Δv = 10 pixel/s). This observation shows consistency with the energy allocation mechanism observed in Atlantic salmon under moderate–low flow conditions [40], confirming common adaptation strategies of aquatic organisms to hydrodynamic environments. In Chen et al.’s (2021) [41] study on juvenile largemouth bass (Micropterus salmoides) in recirculating aquaculture systems, they found that the species showed favorable growth and physiological responses under certain flow conditions. Meanwhile, Li et al. (2019) [20] investigated turbot (Scophthalmus maximus) and reported that they optimized growth at medium flows (0.18 m/s) but exhibited oxidative stress at 0.36 m/s. In contrast, our findings revealed that Litopenaeus vannamei exhibited a distinct ultra-low flow adaptation (0.045 m/s)—a threshold 4–8× lower than that of the aforementioned species. This divergence reflects crustacean-specific energetic economy, where velocity suppression during feeding (Δv = 10 pixel/s) prioritizes digestion over mobility, differing from teleosts’ flow-driven metabolic activation. Such taxon-specific patterns underscore the need for tailored flow management: Teleosts benefit from moderate flow to stimulate metabolism, while crustaceans require sub-critical flows to conserve energy for growth. Notably, the integration of computer vision in our study resolves prior methodological gaps. Unlike discrete biomarker assays (e.g., stress hormones, immune factors) commonly used in fish studies, our framework captures dynamic adaptive behaviors (e.g., real-time velocity adjustments) that directly link flow conditions to ecological fitness, offering a more comprehensive understanding of aquaculture species’ flow responses.

The water circulation rate of 5.0 cycles/day achieved integrated management optimization. The distinct feeding behavior (Δv = 10 pixel/s) synergized with system operational efficiency. Compared to the drawbacks of low flow (2.0 cycles/day: insufficient feed dispersion) and high flow (10.0 cycles/day: movement energy compensation), the core advantages of this parameter are as follows: (1) behavioral identifiability: The velocity difference (Δv = 10 pixel/s) provides a significant signal for detecting feeding states (p < 0.05), facilitating precise feeding control; (2) water quality management: It effectively removes metabolic waste and maintains uniform dissolved oxygen distribution, while avoiding risks associated with lower rates (ammonia accumulation) and higher rates (excessive biofilter scouring); and (3) operational sustainability: It significantly reduces pump operating load compared to 10.0 cycles/day, while the enhanced water quality stability improves shrimp growth efficiency and significantly optimizes the feed conversion ratio (FCR) [42]. This establishes a quantifiable regulatory threshold for recirculating aquaculture systems: When the flow velocity parameter satisfies both behavioral identifiability and energy consumption constraints, it enables the synergistic optimization of precision feeding and energy efficiency.

Compared to existing methods, the MOA framework has distinct strengths and limitations. Its integration of YOLOv8, ByteTrack, and Kalman refraction correction offers advantages: YOLOv8 outperforms YOLOv5 and Faster R-CNN in detecting small, occluded shrimp in turbid water; ByteTrack’s dual-threshold matching reduces ID switches during high-density aggregation (a shortcoming of tools like idtracker.ai); and physics-based refraction correction ensures accurate kinematic parameter extraction (unlike uncorrected vision systems). Limitations include reliance on water transparency (struggling in severe turbidity, unlike multi-modal radar systems), higher computational demands (limiting edge deployment vs. lightweight trackers), and trajectory fragmentation under extreme occlusion (>30 s, unlike ReID-augmented trackers).

Our contributions are threefold: (1) The pipeline enables reliable quantification of shrimp behavior in RAS, capturing feeding/non-feeding velocity differences (e.g., 42.4% reduction at 5.0 cycles/day) for the first time; (2) it provides quantitative evidence for optimizing hydrodynamics, identifying 5.0 cycles/day as an energy-efficient “hydrodynamic window”; and (3) it lays groundwork for ST-GCN-based intelligent feeding models, linking behavior to environmental parameters for precision aquaculture.

5. Conclusions and Outlook

This paper systematically analyzed the network architecture of the ByteTrack algorithm and, through ablation studies and comparative validation, confirmed its performance advantages. Furthermore, based on the industrial recirculating aquaculture system scenario, by integrating multi-dimensional monitoring techniques, the dynamic regulatory mechanisms of water circulation rate on the behavioral characteristics of Litopenaeus vannamei were revealed. By setting up three experimental gradients of water circulation rates (2.0, 5.0, and 10.0 times/day), the impact of water velocity on the behavioral parameters, including movement speed and turning angle, during feeding and non-feeding states of Litopenaeus vannamei was quantitatively analyzed. The core conclusions are as follows: (1) The constructed computer vision analysis framework (YOLOv8-SPD enhanced small object detection + ByteTrack dual-threshold anti-occlusion tracking + Kalman filter correction) can accurately and continuously extract the movement trajectories of shrimp individuals in complex aquaculture environments. (2) The gradient of water circulation rate significantly affects the movement speed of shrimp but has no significant effect on the turning angle. (3) Under only the medium water circulation rate condition (5.0 times/day), the average movement speed of shrimp during feeding states was significantly lower than during non-feeding states (p < 0.05), revealing that the medium flow velocity acts as a “hydrodynamic window period” for inducing feeding behavior-specificity expression. This flow velocity can trigger feeding-priority behavior while avoiding triggering stress responses. (4) Comprehensive analysis indicates that a water circulation rate of 5.0 times/day achieves the optimal balance in inducing significant behavioral differences, maintaining water quality stability, and optimizing energy consumption costs. This study provides an important theoretical basis and practical guidance for the precise regulation of hydrodynamic parameters, energy saving and consumption reduction, and the development of intelligent feeding strategies in industrial recirculating aquaculture systems.

The MOA framework advances behavioral quantification via integrated YOLOv8 detection, dual-threshold ByteTrack tracking, and Kalman-based refraction correction, yet it faces limitations. Technically, its performance declines in turbid water, under surface waves/bubbles (uncorrected spatial errors), and during long-term occlusions (>15 s, irreversible ID switches)—issues worsened by high densities and severe occlusion. Beyond this, the study lacks physiological validation (e.g., cortisol, LDH, glycogen) to clarify whether extreme-flow feeding shifts reflect adaptation or stress, and it underexplores multi-environmental factor interactions (e.g., DO).

Future work will (1) enhance robustness with near-infrared imaging, ByteTrack re-identification integration, and turbulence-adaptive CNN refraction correction; (2) link biomarkers/metabolic indicators to velocity gradients (Figure 4) and explore environmental couplings; and (3) integrate the ST-GCN model with real-time biosensor data to develop closed-loop systems for precise feeding and flow regulation via multi-modal feedback. These steps will strengthen the framework’s practical utility in RAS management.