Vehicle Point Cloud Segmentation Based on Deep Learning

Abstract

A vehicle point cloud segmentation model is designed to address the common issue of real-time acquisition of point cloud of vehicle body and wheels in intelligent parking systems. Firstly, a Multi resolution Feature Pruning Segmentation Network (MFPS-Net) is proposed. This method takes pure point cloud information as input, adopts PointNet++ trunk network, and integrates channel level pruning to accurately segment vehicle point clouds while improving model segmentation efficiency. Then, in order to verify the effectiveness of the model, a dataset with real labels was generated from the vehicle point cloud using bilateral LiDAR data collected from the vehicle buffer. Finally, point cloud segmentation experiments were conducted. The experimental results showed that the accuracy of vehicle point cloud segmentation reached 92.01%.