Robust Fish Enumeration by Multiple Object Tracking in Overhead Videos.

In this paper, we design a framework called detection-tracking fish counting (DTFC) for overhead videos of fish moving in shallow water. DTFC includes three steps, fish tracking, counting zone generation and fish counting. In this work, we adopt two deep learning framework FasterRCNN and DeepSORT as our detector and tracker, respectively. Conventional fish counting methods need to predefine the counting zone, and as a result, they become very inefficient. In DTFC, we use the tracking to refine the generated trajectories and automatically generate the counting zone. For the detection of small objects, we divide the overhead image into four sub-images and enlarge the sub-images for detector training. Our experimental results show that DTFC is very suitable for fish tracking and counting, and may provide a valuable resource for fisheries management.