Deep Learning-Based On-board Laser Point Cloud Road Target Recognition

Lidar is widely used in the fields of autonomous driving, photogrammetry and remote sensing by virtue of its advantages of good concealment, strong anti-interference, high resolution, small size, light weight, and fast acquisition, which can acquire the parameters of target distance, orientation, altitude, speed, attitude and even shape. Aiming at the characteristics of large, sparse and irregular vehiclemounted laser point cloud data, as well as the needs of highprecision measurement of urban roads and automated driving, this paper focuses on the automatic classification and segmentation of vehicle-mounted point cloud data, aiming at accurately identifying the road targets, especially the road edges. To this end, this paper fuses the advantages of PointNet++ and RandLA-Net, and innovatively proposes a multi-scale deformable feature aggregation semantic segmentation network (MSSD-Net). Its core lies in the ShapeDeformCodeModelwithAttention(SDCMA): combining the self-attention mechanism for spatial deformation feature encoding to effectively capture key local Multi-layerattentionup-sampling (MAU): improves the upsampling accuracy and optimises the feature fusion. MSSDNet solves the problem of recognising road edges due to the scarcity of point clouds and the occlusion of vehicles on the roadside.