基于可变形多尺度卷积的三维点云目标检测方法OA

A deformable multi-scale convolutional 3D point cloud target detection method is proposed for the problem of difficult recognition of small targets such as pedestrians.The problem that PointPillars did not consider the contextual information of the point cloud neighborhood was solved by aggregating the corresponding scale point cloud neighborhood information through multilayer convolutional stacking and convolutional kernel setting to enhance the feature expression ability of the network.For the problem that the existing methods neglected the pedestrians' easy deformation characteristics during feature extraction,a two-branching strategy was used to introduce a two-dimensional offset in the shallow feature map,which enabled the network to adaptively learn the weights and improve the robustness to deformation targets.Multi-scale features were extracted using the feature pyramid structure,and the dual-branch features were spliced to obtain fusion information with richer expression.Experiments on the KITTI dataset show that compared with the PointPillars algorithm,this method improves the pedestrian detection accuracy in the bird's eye view mode by 7.83 and 7.38 percentage points for the medium and difficult levels,respectively,and in the 3D mode,the corresponding improvement is 9.91 and 5.42 percentage points.