A two-stage lidar-based approach for enhanced pedestrian and cyclist detection

In recent years, the application scope of LIDAR has been continuously expanding, especially in object detection. Yet existing LIDAR-based methods focus on detecting vehicles on regular roadways. Scenarios with a higher prevalence of pedestrians and cyclists, such as university campuses and leisure centers, have recently received limited attention. To solve this problem, in this paper we propose a novel detection algorithm named SecondRcnn, which is built upon the SECOND algorithm and introduces a novel two-stage detection method. In the first stage, it utilizes 3D sparse convolution on the voxel LIDAR points to learn feature representations. In the second stage, regression is employed to refine the detection bounding boxes generated by the Region Of Interest pooling network. The algorithm was evaluated on the widely used KITTI data set and demonstrated significant performance improvements in detecting pedestrians (4.61% improvement) and cyclist (6.5% improvement) compared to baseline networks. Our work highlights the potential for accurate object detection in scenarios characterized by a higher presence of pedestrians and cyclists. Advancing the use of LIDAR in the field of 3D detection.