基于前景点稠密化的多模态3D目标检测算法OACHSSCD

Aims:Limited by the sampling mechanism of LiDAR,point cloud data usually exhibits significant sparsity,making it difficult to fully describe the geometric structure information of objects.To address this problem,this paper proposes a multimodal 3D object detection method based on foreground point densification.Methods:By fusing image semantic segmentation and depth completion information,targeted enhancement was performed on the point cloud of foreground regions to supplement the missing structural details of objects in the raw point cloud.Meanwhile,a dualbranch parallel network was constructed to balance the completeness of the generated point cloud set and the distribution consistency of the raw point cloud.Results:Extensive experimental results on the KITTI dataset demonstrated that the proposed method achieved consistent performance improvements on three categories of objects:car,pedestrian,and cyclist.It exhibited superior performance especially in challenging scenarios and for medium-to long-range objects.Conclusions:This study combines depth information and semantic information in images,providing a novel solution for 3D object detection tasks.