Scarce fish dataset for aquaculture,an unstable tracking algorithm,and high costs all hinder the technique from being deployed.To solve these problems,a high-performance and real-time fish tracking system was built based on three aspects:fish behavior data,the fish tracking algorithm,and the edge computing platform.In terms of data,the pomfret tracking dataset was collected and labeled on pomfret activity video from the breeding environment.A detection-based fish tracking algorithm named FishTrack was proposed,which enhanced the object association strategy to reduce the tracking lost caused by pose change and long-time occlusion.The FishTrack was deployed on the lightweight edge platform enabling high performance with low computational costs.The results show that FishTrack achieves the tracking accuracy of 72.95%and reduces identification switch of ByteTrack by 83%.Moreover,the tracking system works with a running speed of 7.40 frame/s,which is suitable for real-time tracking under production environment.
关键词
鱼群追踪系统/深度学习/边缘计算/多目标追踪/目标检测
Key words
fish tracking system/deep learning/edge computing/multiple object tracking/object detection
维普
万方数据