Search Results (19)

Real-time roadway surveillance now leans hard on automated detection. How a model trained in one geographic context actually behaves on another, though, is still underexplored for Southeast Asian deployments. We answer that question for Thai roadway closed-circuit television with a cross-domain validation framework. A YOLOv11n (Ultralytics v8.2.0; Ultralytics, Los Angeles, CA, USA) detector trained with focal loss feeds a confidence-aggregation step that turns per-detection scores into a per-frame accident score, and we put four aggregation operators head-to-head. Reliability comes from DeLong variance estimation paired with non-parametric bootstrap on 1245 Thai frames that carry 23 positive accident events. Under maximum-class aggregation the proposed configuration reaches a frame-level AUROC of 0.959 ± 0.020 across three random seeds. Under top-K aggregation it reaches 0.965 ± 0.018. Per-seed DeLong 95 percent intervals exclude chance performance throughout. We also evaluate three baseline configurations: YOLOv5su comes in at 0.738, YOLOv8n at 0.868, and a Chiang Mai-tuned YOLOv11n variant at 0.918. The architectural progression seen on standard benchmarks therefore carries cleanly into the cross-domain setting. The same Chiang Mai-tuned variant reached an in-domain mAP50 of 0.952 yet only 0.918 cross-region AUROC on a separate Thai region, which is a quiet but clear signal that geographic proximity within a country does not on its own remove distributional shift. Bounding-box localisation appears as a secondary diagnostic because the operational target here is frame-level alerting rather than pixel-precise annotation. Edge deployment optimisation falls outside the present scope. What the work leaves behind is a reproducible baseline and a statistical protocol that follow-up Southeast Asian roadway-safety research can build on. Full article

The deployment of object detection models on resource-constrained edge devices remains a substantial challenge, primarily because conventional static networks expend the same worst-case computational cost on every input, regardless of intrinsic difficulty. This paper proposes an input-adaptive dynamic neural network architecture for object detection in embedded environments. The present study investigates two orthogonal axes of input-adaptive inference for embedded object detection: The system demonstrates depth adaptivity through the implementation of Early Exit, and width adaptivity via group-wise Adaptive Routing. The proposed framework is constructed on a frozen Ultralytics YOLO26s backbone and incorporates two YOLO-style early-exit heads positioned at approximately 33% and 66% of the backbone depth. Furthermore, the framework incorporates two Straight-Through Gumbel-Softmax routers, which are appended after Layers 4 and 8 with group-wise hard gating. Both axes additionally accept an explicit external control signal that allows the host system to override the input-conditional policy at inference time. The dual-mode design facilitates the functionality of the trained checkpoint as either an input-adaptive policy, in which the depth and width are determined per sample from the input distribution, or an externally controlled policy. The experimental findings demonstrate two strongly asymmetric input-adaptive policies on a frozen YOLO26s backbone. The early-exit profile reduces the compute per sample from 12.739 to 10.532 GFLOPs—a 17.32% reduction according to our in-house Conv2d/Linear MAC-based GFLOPs estimator—while preserving baseline accuracy (mAP50 = 0.1545 vs. baseline = 0.1528; ΔmAP50 = +0.0017, within evaluation noise; mAP50–95 = −0.0033). Evaluating the router-only profile in the same validator pipeline with a sparsity penalty of γ = 0.05 results in a 12.3% decrease in logical GFLOPs (from 12.739 to 11.172), while maintaining an accuracy level that is at or above the PEFT baseline (mAP50 = 0.2324 and mAP50–95 = 0.1040). In our small-domain PEFT setup, training the dynamic-policy modules yields per-checkpoint mAP shifts in this magnitude. Therefore, we interpret the width-axis accuracy result as preservation of the baseline rather than an improvement. Our contribution on the width axis is reducing computing power while maintaining baseline accuracy. Importantly, the router profile’s logical GFLOP savings are not currently reflected in wall-clock latency under our dense-kernel PyTorch implementation. Achieving practical speedup requires sparse-kernel deployment, such as structured-sparse kernels, TensorRT, TVM, or Triton paths. We will address this in future deployment-level work. Our results indicate that the depth axis can yield genuine end-to-end speedup today, while the width axis offers deployment-pending compute reduction. Full article

Grain number is a primary agronomic trait for targeted yield improvement, with the prospect of enhanced grain production leading to greater food security. Given the complex polygenic nature of the grain number trait, large sample sizes are essential for effective QTL identification. The implementation of trained computer vision models for grain detection offers a timely and cost-effective solution for rapid QTL isolation. In this study, we trained a grain detection model using Ultralytics’ You Only Look Once (YOLOv11) framework. Training was completed on 1000 images of barley spikes, derived from a doubled haploid (DH) population descended from Hindmarsh and RGT Planet. The trained model, termed BarleyGC, achieved satisfactory accuracy metrics (mAP50–95 = 71.9%, recall = 96.7%, precision = 97.1%). Phenotypic characterisation of the DH population was completed with BarleyGC on a distinct collection of 973 images. The Pearson correlation coefficient (r) between model and manual-derived counts for the trait of grain number per spike was 0.895 (p < 0.0001), and 92.4% of all measurements fell within three grains of the manual measurement. Downstream QTL analysis on the phenotype data (n = 153 DH lines), revealed a QTL peak at position 224.959 cM on the genetic map (LOD = 3.14), named qGN-2H. The QTL region contained 21 candidate genes—including HORVU2Hr1G092290 (HORVU.MOREX.r3.2HG0184740), encoding the six-rowed spike 1 (Vrs1) gene—a well-characterised major regulator of row-type divergence and lateral spikelet development. Our study demonstrates the power of the YOLOv11 framework for grain quantification, with BarleyGC capable of grain detection directly from images of intact spikes in two-rowed barley varieties—thus achieving accelerated sample processing for the grain number trait. Full article

Ensuring the structural integrity and operational safety of oil and gas pipelines is a critical challenge due to their extensive geographical coverage and exposure to environmental and anthropogenic risks. Traditional inspection approaches including ground patrols and manned aerial surveys are labor-intensive, costly, and often lack real-time responsiveness. While unmanned aerial vehicles (UAVs) enable flexible and high-resolution monitoring, their practical deployment requires lightweight, robust detection models capable of real-time inference on embedded edge hardware under heterogeneous environmental conditions. This paper presents an end-to-end, edge-deployable UAV inspection framework for simultaneous detection of above-ground pipelines and visually observable anomaly/leak indicators using the official Ultralytics YOLOv26n object detector. A curated dataset of 6127 UAV images acquired across desert, semi-urban, and industrial environments was annotated with two classes (Pipeline and Anomaly/Leak) and partitioned into training 87.5%, validation 8.3%, and testing 4.2% subsets. The detector was fine-tuned from COCO-pretrained weights for 300 epochs at 600 × 600 resolution and evaluated using COCO-style metrics. On the held-out test set, the proposed model achieved 92.4% mAP@0.5 and 75.0% mAP@0.5:0.95, with 89.7% precision, 90.2% recall, and 89.9% F1-score at the selected operating threshold. Optimized TensorRT deployment on an NVIDIA Jetson Xavier NX sustained real-time inference at 18 FPS, demonstrating suitability for onboard UAV processing. Rather than proposing a new detector architecture, the study contributes a domain-specific annotated UAV dataset, deployment-oriented benchmarking, and an end-to-end edge inference workflow for corridor-scale monitoring. The proposed framework can help reduce environmental contamination risk and improve personnel safety during pipeline inspection. Full article

This study addresses the automatic classification of sugar beet seeds according to their spraying levels using RGB images, aiming to enable a fast, practical, and non-destructive early warning system without chemical analysis. A dataset of 16,519 seed images acquired under controlled lighting conditions was used to evaluate YOLOv8-CLS and YOLO11-CLS architectures, including the n, s, m, l, and x scale variants within the Ultralytics framework. All experiments were conducted using a 10-fold cross-validation strategy, with models trained under different batch size and learning rate configurations. The results indicate that both architectures achieve reliable performance, with accuracy values ranging from approximately 78–83% for YOLOv8-CLS and 80–82% for YOLO11-CLS models. ROC-AUC scores consistently above 0.94 demonstrate strong inter-class discrimination. Misclassification analysis shows that errors mainly occur between visually similar intermediate treatment levels, particularly 25% and 50%. Despite this challenge, low log-loss values and balanced precision–recall profiles indicate stable decision behavior. Overall, the findings confirm that sugar beet seed treatment levels can be effectively distinguished using only RGB imagery, providing a potentially low-cost and scalable approach for early warning and quality control in seed treatment processes. Full article

Document Layout Analysis (DLA) is a critical step in intelligent document processing and is essential for accurately reconstructing the hierarchical structure of pages. While modern convolutional neural networks exhibit high performance, their effectiveness heavily depends on the quality and representativeness of training data, limiting their application in scenarios where labeled datasets are scarce. This paper proposes a method for enhancing DLA through synthetic generation of training data. A formalized mathematical model for generating document layouts has been developed, allowing control over element placement density, sizes, and spatial distribution. An experimental study investigated the impact of various data generation strategies on the training of the YOLO11m model, including median and threshold-based element splitting as well as different block sampling schemes. The experiments showed that employing median element splitting combined with random sampling from a large shuffled pool of synthetic data yields consistent improvements of 2–4% across all key metrics: precision, recall, mAP@50, and mAP@50:95, as compared with simple data generation strategies. These results demonstrate that targeted optimization of the data preparation process can enhance the performance of convolutional models in DLA tasks without increasing architectural complexity. The practical applicability of the method is validated through integration into the MinerU system. Future research will focus on extending the proposed model to complex layouts in scientific journals, technical reports, and handwritten documents. Full article

Reliable and up-to-date measurements of public space remain scarce in small and medium-sized towns (SMSTs), where conventional geospatial datasets are often outdated, inconsistent, or inaccessible. This study presents a low-cost and fully reproducible computational pipeline that integrates nadir RGB imagery captured by a DJI Mini 3 UAV with a lightweight instance-segmentation model (Ultralytics YOLOv12-seg) and GIS-based post-processing to derive class-specific surface indicators at the neighborhood scale. The workflow consists of four components: autonomous UAV acquisition over three representative zones of Sesquilé, Colombia; planar mosaic generation and georeferencing using ad hoc ground control points; fine-tuning of a YOLOv12-seg model trained on locally annotated images; and transformation of predicted masks into OSM and GeoPackage geometries for metric analysis. The trained model achieved stable convergence with mask mAP50 ≈ 0.85 and mAP50–95 ≈ 0.70, supported by balanced precision–recall behavior across classes. Spatial outputs exhibit coherent morphological contrasts between the analyzed zones. Buildings occupy 48.17% of the mapped area, vegetation 25.88%, and transport- and plaza-related public space (roadways, sidewalks, and hardscape areas) 25.95%. These proportions capture a clear gradient from a dense urban core to less consolidated peripheral sectors. Results demonstrate that very-high-resolution UAV imagery, combined with open-source deep-learning tools and structured GIS post-processing, can reliably produce operational public-space indicators for SMSTs at low cost. The methodology provides an accessible and scalable framework for evidence-based urban assessment in municipalities with limited technical and financial resources. Full article

Addressing the issue of fruit recognition and localization failures in harvesting robots due to severe occlusion by branches and leaves in complex orchard environments, this paper proposes an occlusion avoidance method that combines a lightweight YOLOv8n model, developed by Ultralytics in the United States, with active perception. Firstly, to meet the stringent real-time requirements of the active perception system, a lightweight YOLOv8n model was developed. This model reduces computational redundancy by incorporating the C2f-FasterBlock module and enhances key feature representation by integrating the SE attention mechanism, significantly improving inference speed while maintaining high detection accuracy. Secondly, an end-to-end active perception model based on ResNet50 and multi-modal fusion was designed. This model can intelligently predict the optimal movement direction for the robotic arm based on the current observation image, actively avoiding occlusions to obtain a more complete field of view. The model was trained using a matrix dataset constructed through the robot’s dynamic exploration in real-world scenarios, achieving a direct mapping from visual perception to motion planning. Experimental results demonstrate that the proposed lightweight YOLOv8n model achieves a mAP of 0.885 in apple detection tasks, a frame rate of 83 FPS, a parameter count reduced to 1,983,068, and a model weight file size reduced to 4.3 MB, significantly outperforming the baseline model. In active perception experiments, the proposed method effectively guided the robotic arm to quickly find observation positions with minimal occlusion, substantially improving the success rate of target recognition and the overall operational efficiency of the system. The current research outcomes provide preliminary technical validation and a feasible exploratory pathway for developing agricultural harvesting robot systems suitable for real-world complex environments. It should be noted that the validation of this study was primarily conducted in controlled environments. Subsequent work still requires large-scale testing in diverse real-world orchard scenarios, as well as further system optimization and performance evaluation in more realistic application settings, which include natural lighting variations, complex weather conditions, and actual occlusion patterns. Full article

To address the challenges of cross-modal feature misalignment and ineffective information fusion caused by the inherent differences in imaging mechanisms, noise statistics, and semantic representations between visible and synthetic aperture radar (SAR) imagery, this paper proposes a multimodal remote sensing object detection method, namely YOLO-CMFM. Built upon the Ultralytics YOLOv11 framework, the proposed approach designs a Cross-Modal Fusion Module (CMFM) that systematically enhances detection accuracy and robustness from the perspectives of modality alignment, feature interaction, and adaptive fusion. Specifically, (1) a Learnable Edge-Guided Attention (LEGA) module is constructed, which leverages a learnable Gaussian saliency prior to achieve edge-oriented cross-modal alignment, effectively mitigating edge-structure mismatches across modalities; (2) a Bidirectional Cross-Attention (BCA) module is developed to enable deep semantic interaction and global contextual aggregation; (3) a Context-Guided Gating (CGG) module is designed to dynamically generate complementary weights based on multimodal source features and global contextual information, thereby achieving adaptive fusion across modalities. Extensive experiments conducted on the OGSOD 1.0 dataset demonstrate that the proposed YOLO-CMFM achieves an $mAP@50$ of 96.2% and an $mAP@50$:95 of 75.1%. While maintaining competitive performance comparable to mainstream approaches at lower IoU thresholds, the proposed method significantly outperforms existing counterparts at high IoU thresholds, highlighting its superior capability in precise object localization. Also, the experimental results on the OSPRC dataset demonstrate that the proposed method can consistently achieve stable gains under different kinds of imaging conditions, including diverse SAR polarizations, spatial resolutions, and cloud occlusion conditions. Moreover, the CMFM can be flexibly integrated into different detection frameworks, which further validates its strong generalization and transferability in multimodal remote sensing object detection tasks. Full article

Reliable differentiation of crops and weeds is essential for precision agriculture, where real-time detection can minimize chemical inputs and support site-specific interventions. This study presents the large-scale and systematic benchmark of 19 YOLO-family variants, spanning YOLOv3 through YOLOv11, for cotton–weed detection using the Cotton–8 dataset. The dataset comprises 4440 annotated field images with five categories: broadleaf weeds, grass weeds, and three growth stages of cotton. All models were trained under a standardized protocol with COCO-pretrained weights, fixed seeds, and Ultralytics implementations to ensure reproducibility and fairness. Inference was conducted with a confidence threshold of 0.25 and a non-maximum suppression (NMS) IoU threshold of 0.45, with test-time augmentation (TTA) disabled. Evaluation employed precision, recall, mAP@0.5, and mAP@0.5:0.95, along with inference latency and parameter counts to capture accuracy–efficiency trade-offs. Results show that larger models, such as YOLO11x, achieved the best detection accuracy (mAP@0.5 = 81.5%), whereas lightweight models like YOLOv8n and YOLOv9t offered the fastest inference ( 27 msper image) but with reduced accuracy. Across classes, cotton growth stages were detected reliably, but broadleaf and grass weeds remained challenging, especially under stricter localization thresholds. These findings highlight that the key bottleneck lies in small-object detection and precise localization rather than architectural design. By providing the first direct comparison across successive YOLO generations for weed detection in cotton, this work offers a practical reference for researchers and practitioners selecting models for real-world, resource-constrained cotton–weed management. Full article

Face detection technology is essential for surveillance and security projects; however, algorithms designed to detect faces in color images often struggle in poor lighting conditions. In this paper, we describe the development of an embedded vision system designed to detect human faces by analyzing images captured with thermal infrared sensors, thereby overcoming the limitations imposed by varying illumination conditions. All variants of the Ultralytics YOLOv8 and YOLO11 models were trained on the Terravic Facial IR database and tested on the Charlotte-ThermalFace database; the YOLO11 model achieved slightly higher performance metrics. We compared the performance of two embedded system boards: the NVIDIA Jetson Orin Nano and the NVIDIA Jetson Xavier NX, while running the trained model in inference mode. The NVIDIA Jetson Orin Nano performed better in terms of inference time. The developed embedded vision system based on these platforms accurately detects faces in thermal images in real-time. Full article

In experimental research, animal welfare should always be of the highest priority. Currently, physical in-person observations are the standard. This is time-consuming, and results are subjective. Video-based machine learning models for monitoring experimental pigs provide a continuous and objective observation method for animal misthriving detection. The aim of this study was to develop and validate a pig tracking technology, using video-based data in a machine learning model to analyze the posture and activity level of experimental pigs living in single-pig pens. A research prototype was created using a microcomputer and a ceiling-mounted camera for live recording based on the obtained images from the experimental facility, and a combined model was created based on the Ultralytics YOLOv8n for object detection trained on the obtained images. As a second step, the Lucas–Kanade sparse optical flow technique for movement detection was applied. The resulting model successfully classified whether individual pigs were lying, standing, or walking. The validation test showed an accuracy of 90.66%, precision of 90.91%, recall of 90.66%, and correlation coefficient of 84.53% compared with observed ground truth. In conclusion, the model demonstrates how machine learning can be used to monitor experimental animals to potentially improve animal welfare. Full article

This article presents a cost-effective camera network system that employs neural network-based object detection and stereo vision to assist a pan–tilt–zoom camera in imaging fast, erratically moving small aerial targets. Compared to traditional radar systems, this approach offers advantages in supporting real-time target differentiation and ease of deployment. Based on the principle of knowledge distillation, a novel data augmentation method is proposed to coordinate the latest open-source pre-trained large models in semantic segmentation, text generation, and image generation tasks to train a BicycleGAN for image enhancement. The resulting dataset is tested on various model structures and backbone sizes of two mainstream object detection frameworks, Ultralytics’ YOLO and MMDetection. Additionally, the algorithm implements and compares two popular object trackers, Bot-SORT and ByteTrack. The experimental proof-of-concept deploys the YOLOv8n model, which achieves an average precision of 82.2% and an inference time of 0.6 ms. Alternatively, the YOLO11x model maximises average precision at 86.7% while maintaining an inference time of 9.3 ms without bottlenecking subsequent processes. Stereo vision achieves accuracy within a median error of 90 mm following a drone flying over 1 m/s in an 8 m $\times$ 4 m area of interest. Stable single-object tracking with the PTZ camera is successful at 15 fps with an accuracy of 92.58%. Full article

Estimating the calorie content of fruits is critical for weight management and maintaining overall health as well as aiding individuals in making informed dietary choices. Accurate knowledge of fruit calorie content assists in crafting personalized nutrition plans and preventing obesity and associated health issues. In this paper, we investigate the application of deep learning models for estimating the calorie content in fruits from digital images, aiming to provide a more efficient and accurate method for nutritional analysis. We create a dataset comprising images of various fruits and employ random data augmentation techniques during training to enhance model robustness. We utilize the RT-DETR model integrated into the ultralytics framework for implementation and conduct comparative experiments with YOLOv10 on the dataset. Our results show that the RT-DETR model achieved a precision rate of 99.01% and mAP50-95 of 94.45% in fruit detection from digital images, outperforming YOLOv10 in terms of F1- Confidence Curves, P-R curves, precision, and mAP. Conclusively, in this paper, we utilize a transformer architecture to detect fruits and estimate their calorie and nutritional content. The results of the experiments provide a technical reference for more accurately monitoring an individual’s dietary intake by estimating the calorie content of fruits. Full article

This study introduces a You Only Look Once (YOLO) model for detecting diseases in tomato leaves, utilizing YOLOV8s as the underlying framework. The tomato leaf images, both healthy and diseased, were obtained from the Plant Village dataset. These images were then enhanced, implemented, and trained using YOLOV8s using the Ultralytics Hub. The Ultralytics Hub provides an optimal setting for training YOLOV8 and YOLOV5 models. The YAML file was carefully programmed to identify sick leaves. The results of the detection demonstrate the resilience and efficiency of the YOLOV8s model in accurately recognizing unhealthy tomato leaves, surpassing the performance of both the YOLOV5 and Faster R-CNN models. The results indicate that YOLOV8s attained the highest mean average precision (mAP) of 92.5%, surpassing YOLOV5’s 89.1% and Faster R-CNN’s 77.5%. In addition, the YOLOV8s model is considerably smaller and demonstrates a significantly faster inference speed. The YOLOV8s model has a significantly superior frame rate, reaching 121.5 FPS, in contrast to YOLOV5’s 102.7 FPS and Faster R-CNN’s 11 FPS. This illustrates the lack of real-time detection capability in Faster R-CNN, whereas YOLOV5 is comparatively less efficient than YOLOV8s in meeting these needs. Overall, the results demonstrate that the YOLOV8s model is more efficient than the other models examined in this study for object detection. Full article