FP-RCNN: A Real-Time 3D Target Detection Model based on Multiple Foreground Point Sampling for Autonomous Driving

The perception module of autonomous driving needs to maintain high detection accuracy and speed in various weather conditions. Two-dimensional target detection maintains fast detection speed but poor detection accuracy in bad weather, whereas three-dimensional (3D) target detection can still have a good detection effect in bad weather. However, in the current 3D target detection methods, the accuracy of single-stage detection algorithms is insufficient, and the speed of two-stage detection algorithms is slow. Therefore, in this study, we propose a real-time 3D target detection model based on multiple foreground point sampling for autonomous driving, FP-RCNN. FP-RCNN incorporates features from the original points, voxels, and birds-eye view (BEV), uses sparse convolution at the voxel level, performs multiple progressive downsampling to extract features, and maps the obtained features on BEV to obtain BEV features. A three-layer progressive sampling structure was used for key point sampling. The third layer uses instance-aware downsampling to exploit semantic information to ensure that as many foreground points as possible are collected, and the three features are subjected to VSA operations to obtain the final features bound to sampled key points. The second stage divides the proposed box obtained in the first stage, fuses the contextual information of the original points to obtain the final point features, and outputs the confidence box through two fully connected layers. FP-RCNN is tested on the KITTI dataset, and the test results show a 6% improvement in pedestrian detection and a 50% improvement in detection speed compared with a representative two-stage approach.