IEEE Access (Jan 2024)

A Measurement Model for Aquatic Animals Based on Instance Segmentation and 3D Point Cloud

  • Zhiqian He,
  • Xiaoqing Xu,
  • Jialu Luo,
  • Ziwen Chen,
  • Weibo Song,
  • Lijie Cao,
  • Zhongming Huo

DOI
https://doi.org/10.1109/ACCESS.2024.3470791
Journal volume & issue
Vol. 12
pp. 156208 – 156223

Abstract

Read online

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. Most measurement models for the size of aquatic animals measure length and width. Therefore, this paper proposes a Point Cloud Measurement Model to overcome the problems caused by the image scale change and feature extraction difficulty in aquatic animal measurement models. The proposed model comprises an image instance segmentation YOLOv8 model, a 3D point cloud, and a depth camera. First, the improved YOLOv8 performs instance segmentation on the aquatic animal to output Mask and Box coordinates. Then, the 3D point cloud of the aquatic animal is reconstructed based on the Mask and Box combined with the depth camera. Secondly, the Mask is processed to obtain the pixel end point coordinates of the length and width, and the actual length and width information is obtained through the depth value. Finally, Box was used to fitting the point cloud planes while obtaining the heights of aquatic animals from the point cloud planes. To evaluate the effectiveness of the improved instance segmentation model YOLOv8, we used the self-made Aquatic Animal SegMentation dataset(AASM), the Underwater Image Instance Segmentation (UIIS) dataset, and the Underwater Salient Instance Segmentation (USIS10K) Dataset for verification. The experimental results show that the improved YOLOv8 model compared with the baseline model improves the [email protected] by 97.5% on the self-made underwater instance segmentation dataset, the [email protected] by 3.5% on the UIIS dataset, and the [email protected] by 1.5% on the USIS10K dataset. At the same time, the proposed point cloud measurement model has an absolute error of less than 5 mm in length, width, and height compared to manual measurement. The absolute height error of the clam is 0.89 mm. The experimental results demonstrate the versatility and accuracy of the proposed point cloud measurement model.

Keywords