Abstract: Segmenting the point cloud of an object from the point cloud data of underwater environment is a key step in underwater target recognition and reconstruction. In order to segment the point cloud of the object from the underwater 3D point cloud data more accurately, the traditional region growing 3D point cloud segmentation algorithm is improved. In the procedure of preprocessing the point cloud data, noise data such as invalid points and outlier points are removed by straight-through filtering, statistical filtering and radius filtering, and next the point cloud data is simplified by using voxel down-sampling, afterwards the point cloud is roughly segmented by using a density-based spatial clustering of applications with noise(DBSCAN) clustering algorithm, thus the area where the number of point cloud points exceeds a set threshold are clustered, and the boundary points are further determined by using an angle criterion algorithm as the main basis for selecting regional growth seed points. In this way, the over-segmentation or under-segmentation that are usually the side effects of the traditional region growing algorithm are alleviated, so as to achieve more accurate segmentation of the point cloud. Two typical segmentation algorithms, namely the Euclidean clustering and single DBSCAN, are used to compare with the proposed method. Through a set of experiments conducted on the measured data, the segmentation accuracy, recall and F_1 value of point cloud data in an experimental pool are 98.6%, 83.5% and 90. 8%, respectively, which show significant improvements compared with the traditional region growing algorithm, and the results also show that the proposed method manifests higher accuracy in contrast with the other two algorithms. The method proposed in this paper demonstrates good performance for the underwater environment point cloud data surveyed by 3D sonar.