Scholarlink

<p>Manual counting of rice field eel (<i>Monopterus albus</i>) during high-throughput operations, such as tank transfer and grading, is labor-intensive and prone to error. In chute scenarios, dense and high-speed downstream eel flows are frequently accompanied by body entanglement and occlusion, which makes reliable vision-based counting challenging. This study proposes EelTrack-Edge, an edge-deployable detection–tracking–counting framework for high-throughput rice field eel counting. To reduce ambiguity caused by slender-body overlap, eel heads are used as the unified targets for detection, tracking, and counting, and a high-frame-rate annotated dataset is constructed accordingly. For detection, a lightweight YOLO11-based model is developed, reducing parameters and FLOPs by 37% and 46%, respectively, while achieving an mAP50 of 92.86%. For tracking, ByteTrack is improved by reformulating the Kalman-filter state representation and introducing a flow-aligned motion-corridor constraint. Under high-density conditions, the improved tracker achieves a MOTA of 76.8% and an IDF1 of 85.8%. For counting, an oriented virtual counting line with a dual-threshold hysteresis mechanism is designed to suppress duplicate counts, reaching an average counting precision (ACP) of 97.06%. Frame-rate ablation shows that reduced temporal resolution increases inter-frame displacement, weakens association stability, and leads to missed or duplicate counts. On an NVIDIA Jetson AGX Orin with TensorRT FP16 acceleration, the YOLO11s-Eel detector achieves an average single-frame inference latency of approximately 4.8 ms, and the complete pipeline reaches 122 FPS at 640 × 640 resolution while maintaining an ACP of 95.86%. These results demonstrate that EelTrack-Edge provides an accurate and real-time solution for low-power edge deployment in high-throughput rice field eel counting.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

High-throughput rice field eel counting: an edge-deployable method via lightweight tracking

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Yuntao Zhai,
  • Chunyan Zhang,
  • Quan Yuan,
  • Hang Yang,
  • Zhen Xu,
  • Wenzong Zhou

摘要

Manual counting of rice field eel (Monopterus albus) during high-throughput operations, such as tank transfer and grading, is labor-intensive and prone to error. In chute scenarios, dense and high-speed downstream eel flows are frequently accompanied by body entanglement and occlusion, which makes reliable vision-based counting challenging. This study proposes EelTrack-Edge, an edge-deployable detection–tracking–counting framework for high-throughput rice field eel counting. To reduce ambiguity caused by slender-body overlap, eel heads are used as the unified targets for detection, tracking, and counting, and a high-frame-rate annotated dataset is constructed accordingly. For detection, a lightweight YOLO11-based model is developed, reducing parameters and FLOPs by 37% and 46%, respectively, while achieving an mAP50 of 92.86%. For tracking, ByteTrack is improved by reformulating the Kalman-filter state representation and introducing a flow-aligned motion-corridor constraint. Under high-density conditions, the improved tracker achieves a MOTA of 76.8% and an IDF1 of 85.8%. For counting, an oriented virtual counting line with a dual-threshold hysteresis mechanism is designed to suppress duplicate counts, reaching an average counting precision (ACP) of 97.06%. Frame-rate ablation shows that reduced temporal resolution increases inter-frame displacement, weakens association stability, and leads to missed or duplicate counts. On an NVIDIA Jetson AGX Orin with TensorRT FP16 acceleration, the YOLO11s-Eel detector achieves an average single-frame inference latency of approximately 4.8 ms, and the complete pipeline reaches 122 FPS at 640 × 640 resolution while maintaining an ACP of 95.86%. These results demonstrate that EelTrack-Edge provides an accurate and real-time solution for low-power edge deployment in high-throughput rice field eel counting.