A 3D Point Cloud Object Detection Algorithm Based on MSCS-Pointpillars

As an important technique of driverless vehicles, 3D Point Cloud Object Detection algorithm can provide semantic information and geometrical information of various types of objects. For the problem of feature information loss in 3D Point Cloud Object Detection, this paper proposes an improved Pointpillars algorithm where a multi-scale columnar feature extraction network based on the attention mechanism. In this algorithm we extract features to obtain pseudo-images of different scales, and splicing point cloud pseudo-images of multiple scales to obtain fusion feature maps. In addition, by introducing a Convolutional Block Attention Module (CBAM), our MSCS-Piontpillars can effectively suppress the noise in the pseudo-image of the point cloud and amplify the important feature information for target classification. Comparision experimental was carried on both the KITTI 3D Object dataset and a real experiment. The results show that compared with Pointpillars, MSCS-Pointpillars (Multi-Scale Channel Spatial Attention Pointpillars) have improved the detection accuracy of cars, cyclists, pedestrians and other targets.