Voxel-RCNN Improved Algorithm Based on Atrous Spatial Pyramid Pooling

In order to solve the problem of poor bicycle and pedestrian detection under sparse point cloud conditions in autonomous driving, an improved Voxel-RCNN algorithm is proposed to optimize the performance by introducing a dilated spatial pyramid structure while minimizing the computational cost and model complexity. Voxel-RCNN uses 3D sparse convolution with a fixed-size kernel to extract target features, which may not be suitable for the detection of objects of different sizes and limits the information interaction of long-distance features. To this end, we design a sparse atrous spatial pyramid pooling (SASPP) module to expand the receptive field of the model and fuse multi-scale features. For the same purpose, a ASPP module is integrated in the 2D backbone network. The experimental results show that the improved algorithm has achieved significant improvement in the detection of bicycles and pedestrians, especially on difficult samples. The pedestrian detection accuracy is improved by 4.65%, and the bicycle detection accuracy is improved by 4.23%. These results demonstrate that our algorithm is effective for difficult-to-detect objects such as bicycles and pedestrians.