Multi-layer Pointpillars: Multi-layer Feature Abstraction for Object Detection from Point Cloud

In order to extract the spatial structure features of the original point cloud, multi layers pointpillars model, a fast and efficient one-stage network, is proposed for object detection from point cloud. Firstly, point cloud are divided into multi layers along z axis, by each layer to generate pillars in the vertical direction, and multi layers pseudo-image representing for multi layers are created by the method of pointpillars. Then, the multi layers and complete pseudo-image are fused as the input of RPN, and the feature maps with context information and multi-scale features are obtained. Finally, the detection boxes and classification score were obtained by SSD head according to the feature maps. We get a high quality prediction box and classification results. The experimental results show that multi-layer pointpillars can get higher precision than the original pointpillars.