Method, device, computer system for detecting pedestrian based on 3d point clouds

A device for detecting pedestrians based on 3D point clouds, comprising:

• a sensor configured to obtain spatial radar point cloud data of area to be detected, wherein the spatial radar point cloud data includes a plurality of radar point cloud data;
• a processor, and a plurality of modules executed by the processor, the plurality of modules comprising:
• a voxel partitioning module, configured to divide the spatial radar point cloud data to obtain a plurality of 3D voxel grid cells according to a preset unit of voxel, wherein the plurality of 3D voxel grid cells comprises a plurality of radar point cloud data;
• a voxel encoding module, configured to encode the plurality of 3D voxel grid cells and obtain the voxel encoded data of the plurality of radar point cloud data corresponding to the plurality of 3D voxel grid cells;
• a sparse convolutional backbone network module, configured to input the voxel encoded data of the plurality of radar point cloud data corresponding to the plurality of 3D voxel grid cells into a predetermined sparse convolutional backbone network for feature extraction to obtain a first feature map;
• a self-attention transformation network module, configured to divide the plurality of 3D voxel grid cells respectively according to a predetermined window to obtain a plurality of 3D sub-voxel grid cells corresponding to the plurality of 3D voxel grid cells and to obtain voxel encoded data of the plurality of radar point cloud data corresponding to the plurality of 3D sub-voxel grid cells;
• and to obtain a second feature map according to the voxel encoded data of the plurality of radar point cloud data corresponding to the plurality of 3D sub-voxel grid cells and a predetermined second feature extraction algorithm, wherein the second feature extraction algorithm is:
• F2=W2(LN(MSA(LN(F1),PE(I))+F1))+b2+MSA(LN(F1),PE(I))+F1,
• wherein F1 represents the voxel encoded data, F2 represents the second feature map, MSA() represents a multi-headed self-attention function, LN(□) represents a layer normalization function, PE() represents a position encoding function, W2 represents a second trainable weight parameter, b2 represents a second bias parameter, and I represents coordinate data of the radar point cloud data corresponding to the plurality of 3D sub-voxel grid cells on the first feature map;
• a pedestrian detection module, configured to fuse the first feature map and the second feature map to obtain a fused feature map, and to input the fused feature map into a predetermined pedestrian detection model for pedestrian detection to obtain the pedestrian detection information for the area to be detected.