Fish Trajectory Extraction Based on Object Detection

Extracting fish trajectories from underwater videos is an important way to analyze fish behavior and can provide guidance for aquaculture. Most of the existing researches are based on traditional image processing algorithms and CamShift tracking algorithm. In recent years, convolutional neural networks have performed better than traditional image processing algorithms in many computer vision tasks. This paper proposes a fish trajectory extraction method based on object detection. First, the deep learning object detection model Faster RCNN is used to predict fish for all video frames. Then based on the criteria of IoU, center distance and probability similarity, a greedy algorithm or Hungarian algorithm is used to correlate the prediction results. Finally, improvements are done from two aspects of linking and deleting. Experiments show that this method can well complete the task of fish trajectory extraction, and the AP performance is increased from 74.75% to 80.94%.