2024 Volume 5 Issue 2 Pages 13-21
Environmental impact assessments and natural environment surveys associated with development projects require, among other tasks, surveys of the habitats of fish and other species. These fish surveys generally involve on-site capture surveys and visual identification. However, it has been observed that stress from such capture surveys has an adverse effect on the fish. Furthermore, such surveys require significant human effort, including the task of visual identification by technical experts. Therefore, as the need for an efficient automatic fish species recognition and counting system is needed. We built a system based on the YOLOv7 deep-learning object detector and a tracking algorithm that automatically distinguishes and counts multiple fish species in underwater video footage. After verifying the performance of the recognition model, we evaluated the system’s accuracy in counting each fish species. The evaluation showed an accuracy rate of 94.1%, demonstrating highly accurate automatic recognition even in a pond environment with low transparency.
