FishMOT: A Simple and Effective Method for Fish Tracking Based on IoU Matching
Shuo Liu,Lulu Han,Xiaoyang Liu,Junli Ren,Fang Wang,YingLiu,Yuanshan Lin
DOI: https://doi.org/10.48550/arXiv.2309.02975
2023-09-06
Computer Vision and Pattern Recognition
Abstract:The tracking of various fish species plays a profoundly significant role in understanding the behavior of individual fish and their groups. Present tracking methods suffer from issues of low accuracy or poor robustness. In order to address these concerns, this paper proposes a novel tracking approach, named FishMOT (Fish Multiple Object Tracking). This method combines object detection techniques with the IoU matching algorithm, thereby achieving efficient, precise, and robust fish detection and tracking. Diverging from other approaches, this method eliminates the need for multiple feature extractions and identity assignments for each individual, instead directly utilizing the output results of the detector for tracking, thereby significantly reducing computational time and storage space. Furthermore, this method imposes minimal requirements on factors such as video quality and variations in individual appearance. As long as the detector can accurately locate and identify fish, effective tracking can be achieved. This approach enhances robustness and generalizability. Moreover, the algorithm employed in this method addresses the issue of missed detections without relying on complex feature matching or graph optimization algorithms. This contributes to improved accuracy and reliability. Experimental trials were conducted in the open-source video dataset provided by idtracker.ai, and comparisons were made with state-of-the-art detector-based multi-object tracking methods. Additionally, comparisons were made with idtracker.ai and TRex, two tools that demonstrate exceptional performance in the field of animal tracking. The experimental results demonstrate that the proposed method outperforms other approaches in various evaluation metrics, exhibiting faster speed and lower memory requirements. The source codes and pre-trained models are available at: https://github.com/gakkistar/FishMOT