High-Throughput Rice Field Eel Counting: An Edge-Deployable Method via Lightweight Tracking

Manual counting of rice field eel ( Monopterus albus ) during high-throughput operations (e.g., tank transfer and grading) is labor-intensive and error-prone. In chute scenarios characterized by high-speed, dense downstream eel flows with frequent entanglement and occlusion, we propose EelTrack-Edge, an edge-deployable detection–tracking–counting framework. To mitigate failures caused by slender-body entanglement, eel heads are defined as the unified targets for detection, tracking and counting, and a high-frame-rate annotated dataset is constructed accordingly. On the detection side, a lightweight YOLO11-based model reduces parameters and FLOPs by 37% and 46%, respectively, while achieving an mAP50 of 92.86%. ByteTrack was improved in two ways: the Kalman-filter state was reformulated, and a flow-aligned motion-corridor constraint was introduced. Under high-density conditions, the improved tracker achieved a MOTA of 76.8% and an IDF1 of 85.8%. For counting, an oriented virtual counting line with a dual-threshold hysteresis mechanism suppresses double counting, reaching an average counting precision (ACP) of 97.06%. Frame-rate ablation further indicates that reduced temporal resolution increases inter-frame displacement and degrades association stability, which propagates to missed and duplicate counts. On an NVIDIA Jetson AGX Orin with TensorRT FP16, the deployed system achieved a latency of approximately 4.8 ms and a throughput of 122 FPS at an input resolution of 640 × 640. The ACP remained at 95.86%, indicating the feasibility of low-power edge deployment.