Abstract
Traditional computer vision measurement methods often encounter challenges such as scale and dynamic changes and feature extraction difficulties when confronted with aquatic animals images, making measuring aquatic animals' morphology complex and restricted. As a result, most of the current models for measuring aquatic animals’ dimensions focus on length and width information. This paper establishes a Point Cloud Measurement Model to overcome the problems caused by image scale changes and difficulty in feature extraction in aquatic animals’ measurement models. The proposed model integrates neural network instance segmentation, 3D point cloud, and depth camera. First, a convolutional neural network is employed to extract and segment the features of aquatic animals to output Mask and Box coordinates, which can be employed to acquire the depth value of the identified aquatic animals. Then, the Point Cloud Measurement Model reconstructs the 3D point cloud of the aquatic animals and processes the Mask. Finally, the height of the vertex of the aquatic animals from the plane is determined by fitting the point cloud to the plane, and the edge detection is applied to the Mask to obtain the length and width, thus acquiring a collection of boundary points for processing. The self-produced aquatic animals’ segmentation dataset and the public Underwater Robotics Professional Competition (URPC) dataset were tested in different network models to evaluate the effectiveness of the proposed model. The experimental results indicate that the mAP@0.5:0.95 of the improved YOLOv8n model is 97.5% on the self-produced dataset, while the mAP@0.5 is 84.6% on the URPC dataset. The absolute errors of length, width, and height are all within 5 mm. The clams’ absolute height error is 0.89 mm. These results demonstrate the generality and accuracy of the proposed point cloud measurement model.