Feature Point Recognition of LiDAR Measurement Data Based on Mask R-CNN

Directly using LiDAR measurement data to extract key information for feature point recognition cannot directly distinguish whether points belong to the same target. Only extracting local feature points can lead to a decrease in data feature recognition accuracy. A LiDAR measurement data feature point recognition based on Mask R-CNN is proposed. Firstly, PointNet++is selected as the backbone network of Mask R-CNN to extract feature vectors, and a feature pyramid network is constructed next to the backbone branch to extract multi-scale features. A three-dimensional candidate box is generated by the region suggestion network, which is input into the classifier network through ROI Align. Target category prediction, candidate box position regression, and binary mask are carried out to output the target segmentation results. Then, based on the segmented target point cloud, a 4D Shepard surface estimation is used. Target point cloud curvature, Obtain the volume integral invariants and normalize them. Finally, cluster the volume integral invariants using the K-means algorithm to achieve feature point recognition in LiDAR measurement data. The experimental results show that the proposed method can effectively segment targets in LiDAR measurement data with a simplification rate of 37.68%. The data feature point recognition performance and quality are high, and the detection results of AP, AP₅₀, and AP₇₅ are all above 90%, indicating good application effects.