A Fish Detection and Tracking Method Based on Improved Interframe Difference and YOLO-CTS

Aiming at the problems of misjudgment, missed judgment, and difficult identification of fish in complex fishway environments, the fish detection method based on improved frame difference and you only look once-CBAM-Transformer-SIOU (YOLO-CTS) is proposed, which is tracked by deep-sort algorithm. First, the existence situation of fish is determined through processing by interframe difference, dilation, and erosion. Then, by adding convolutional block attention module (CBAM) and Transformer modules the feature extraction capability of the model is improved on the basis of the YOLOv5s model. Afterward, SIOU loss is used to recalculate the loss function using the vector angle in the bounding box. The average accuracy (mAP) of the model reaches 98.2%, which is 0.22% higher than YOLOv8s. Finally, the deep-sort tracking algorithm is used to trace the fish. The improved interframe difference method can effectively reduce the misjudgment of abnormal conditions, such as water bubbles and sediment. The self-built fish video collected from the field experiment is used to verify the feasibility and stability of the method. Compared with the existing fish detection algorithms based on traditional image and deep learning, the accuracy of the proposed method reaches 98.3%.