SWFNet: Efficient Sampling and Weighted Fusion for Enhanced PointPillars in 3D Object Detection

3D object detection based on LiDAR point cloud is one of the key techniques for autonomous driving applications. However, it is not easy to extract effective features from complicated point cloud data and keep a balance between accuracy and speed, because the density of the point cloud data changes over time, usually showing different structures in spatial domain, such as near-dense, far-sparse and empty. In this work, we propose an Efficient Sampling Weighted Fusion encoder (SWFNet) for 3D object detection. First, an efficient point cloud sampling method is employed to select valid points for each pillar, which reduces the computational complexity and improves the detection accuracy to some extent. In addition, a pillar feature learning network, called WFNet herein, is designed to extract key point information. It uses the weighted fusion technique, a finer-grained pillar feature, including three dimensions, i.e., point, feature and pillar, can be obtained. Furthermore, by replacing the encoder part with the SWFNet in the PointPillars framework, the detection performance can be improved significantly, especially for detecting small objects. The superior performance of the proposed SWFNet has been demonstrated using the KITTI dataset, showing that the 3D mean average precision (mAP) has improved by 4.33% and the inference speed has been increased up to 115 frames per second (FPS), respectively.