Fish Identification and Tracking Based on Pose Estimation

Traditionally, fish farming involves manual methods for fish identification. However, with the advancement of computer vision technology, an automatic identification system can be developed. This study introduces a fish identification method based on detecting fish body pose using a lightweight convolutional neural network model. The Micropterus salmoides was used as the study object to construct a fish pose dataset. The target detection frame has also been updated to a rotating format to better recognize fish in all orientations. Feature vectors were constructed based on the pose estimation, and the random forest algorithm was applied for feature degradation and fish identity identification. The fish pose detection network achieved an accuracy rate of 0.98 on the test set. Detection-based tracking was also used to track the trajectory of fish targets. This method can provide a foundation for fish individual identification, size detection, and precision breeding devices, optimizing efficiency in fish farming.