Selective Multi-Scale Feature Aggregation Framework for 3D Road Object Detection

In intelligent transportation, 3D object detection is a key task for autonomous vehicles. It is necessary to consider the multi-scale features aggregation due to the complexity of the point cloud. Based on the perspective, a novel selective multi-scale feature aggregation framework is proposed, which has two modified backbone based on the OpenPCDet framework: Multi-Weights 2D Backbone and Feature Separation 3D Backbone. The former module contains a progressive fusion structure considering the characteristics of the Bird's Eye View (BEV) map in multi-category object detection, and the latter module processes and fuses the location information and the reflection intensity information separately based on the difference between them in raw point cloud. The experiments on KITTI dataset show the progress in multi-category object detection while achieving superior performance on the most common Car category, which reflects the effectiveness of the work.