3D Object Detection with Fusion of Max-Sum Encoding and Gaussian Filtering

Environmental perception plays a key role in ensuring the operational safety of self-driving vehicles. It involves sensing spatial information such as the size, location, orientation of car, pedestrian and cyclist in the surrounding environment. Detection algorithms that fuse LiDAR point clouds and images can effectively improve the sparsity defect of point cloud data and the sensitivity of image data to harsh environment, but most fusion detection algorithms have low detection accuracy for small objects such as pedestrian and cyclist. To enhance the perception of small objects, we propose a novel 3D object detection model that integrates Max-Sum encoding and Gaussian filtering into the existing F-PointNet method. By incorporating the Max-Sum encoding module, we effectively aggregate the local fine-grained features and the maximum response information of the point cloud, solving the problem that the F-PointNet network only retains the maximum response but lacks local fine-grained features; we propose a Gaussian distance filtering based instance segmentation strategy to calculate the Gaussian distance between the point cloud and the center of the 2D object, and incorporate the Gaussian distance feature into the instance segmentation, improving the problem of low accuracy of instance segmentation in F-PointNet networks. To validate our algorithm, we conducted experiments in the KITTI and DAIR-V2X datasets. The results show that our algorithm outperforms the F-PointNet benchmark method in general in terms of car, pedestrian and cyclist detection results. Notably, it shows the most significant improvement for small objects such as pedestrian and cyclist, with improvements of 1.47% and 3.87% at the KITTI medium difficulty level, 4.44% and 2.06% at the DAIR-V2X medium difficulty level. In summary, our proposed algorithm effectively solves the low accuracy problem of instance segmentation and small objects detection in F-PointNet networks, and has strong applicability.