Research on detection algorithms for occluded vehicles on urban roads
To enhance the detection accuracy of occluded vehicles on urban roads for intelligent vehicles,this paper proposes a bi-scale point cloud density expansion network (BPDE-Net) to address the sparse object point cloud issue caused by vehicle occlusion.First,the original point cloud is projected onto image labels after semantic segmentation,and a fixed number of virtual points are randomly generated within occluded areas.A mixed interpolation method is employed to associate virtual points with actual projected points,and the obtained virtual points are then reverse-mapped to the point cloud space.Next,the Mahalanobis distance is utilized to associate point clouds between adjacent voxels,increasing the number of similar point clouds within each voxel.An improved attention Gaussian matrix is used to calculate the relative position encoding corresponding to voxel features,focusing on the relative positions of different voxel sequences within the channel.Extensive comparative experiments are conducted on urban roads with occlusions between numerous vehicles selected from the KITTI dataset.Our results indicate the average detection accuracy of occluded vehicles under both 3D and bird's-eye-view perspectives by BPDE-Net reaches 79.2％ and 83.7％ respectively.Compared to the baseline network Second,BPDE-Net improves by 11.2％ and 12.5％ respectively.Additionally,the mAP of the point cloud density enhancement module and voxel feature fusion module is increased by 3.9％ and 1.8％ respectively compared to current mainstream methods,enhancing the recognizability and robustness of vehicles in occluded scenarios.
obstructed vehicle detectionattention mechanismpoint cloud density enhancementvoxel features fusionmultimodal fusion
国家科技期刊平台
NSTL
万方数据
