CP-RCNN: Lidar Object Detection with Feature Pooling and Abstraction

LiDAR-based object detection is a challenging task for autonomous navigation systems, especially in pedestrian-rich environments. Recently, the integration of deep learning techniques with lidar-generated point cloud data has advanced object detection and segmentation in many scenarios. However, current lidar-based methods usually struggle to accurately detect small-sized objects, such as pedestrian and cyclist, causing severe safety and reliability concerns for autonomous vehicles. This study refines structural design of lidar based neural networks to enhance precision and recall metrics for the identification of small entities. Specifically, we introduce CP-RCNN, a novel lidar object detection framework that combines state of the art voxelization and feature extraction techniques. Extensive ablation experiments demonstrate that our method has improved performance in the detection of pedestrians and cyclists. Furthermore, this paper also proposes a novel neural network structure named Centerpoint-RCNN, which not only maintains high precision in vehicle classification but also achieves an impressive inference speed of 15Hz on the NVIDIA RTX 4090 graphics processing unit.