Abstract: 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 point cloud neighborhood was solved by aggregating corresponding scale point cloud neighborhood information through multilayer convolutional stacking and convolution kernel setting to enhance feature expression ability. For the problem that existing methods neglected pedestrians’ easy deformation characteristics during feature extraction, a two-branching strategy was used to introduce two-dimensional offset in shallow feature map, which enabled the network to adaptively learn weights and improve robustness to deformation targets. Multi-scale features were extracted using feature pyramid structure, and dual-branch features were spliced to obtain fusion information with richer expression. Experiments on the KITTI dataset show that compared with PointPillars algorithm, this method improves pedestrian detection accuracy by 7.83 and 7.38 percentage points for medium and difficult levels in bird’s eye view mode, and the corresponding improvement is 9.91 and 5.42 percentage points in 3D mode.