A Binocular Vision Fish Body Length Measurement Method Based on Improved YOLOv8-pose

The size information of farmed fish reflects their developmental status. To accurately obtain the body length information of fish in different postures, a binocular vision fish body length measurement method based on improved YOLOv8 pose is proposed. Firstly, a keypoint detection model called MultiScale_CBAM_YOLO (MSC_YOLO) is proposed by introducing an improved MS_StemBlock module at the forefront of the original YOLOv8-pose network's Backbone and incorporating convolutional block attention mechanism (CBAM) into the network. Secondly, the MSC_YOLO model is used to detect the key point coordinates of the tested fish, and combined with the binocular imaging model, the three-dimensional coordinate reconstruction of the key points is completed. Finally, generate a smooth fish pose curve to calculate the body length of the fish. The experimental results showed that the mAP50, mAP50-95, and precision of the MSC_YOLO model were 88.84%, 87.39%, and 92.31%, respectively. When the imaging angles are 0°~20 °, 20 °~40°. and 40°~60°, the average absolute errors of fish body length measurement results compared to the true values are 0.23cm, 0.46cm, and 0.70cm, respectively, and the average absolute percentage errors are 1.47%, 2.90%, and 4.40%, respectively. The research results can provide technical support for measuring the body length of farmed fish.