Fish Length Estimation Based on Stereo Vision and Keypoint Detection

Fish length information is of great significance in aquaculture. However, traditional fish length measurement methods are not only time-consuming and labor-intensive but can also impact fish growth. Therefore, this paper proposes an automatic non-contact fish length measurement system based on stereo vision and keypoint detection. The process begins with the calibration of the stereo images. The rectified left image is input into the YOLOv7 network to detect fish suitable for length measurement. The detected fish are then passed through an improved HRNet (FPN-HRNet) to identify key points on the fish in the image. Using the coordinates of these key points and corresponding disparity values, 3D spatial coordinates are calculated. To reduce the occurrence of significant errors in fish length measurements due to inaccuracies in the 3D coordinates of keypoints, a curve fitting is applied to these spatial coordinate points. After fitting, the spatial distances between these points are calculated to obtain the fish length. The experimental results show that the average relative error of length measurement is 3.01%, which indicates that the proposed fish length measurement system can accurately and reliably measure the length of fish.