LiDAR-PDP: A LiDAR-Based Panoptic Dynamic Driving Environment Perception Algorithm

A LiDAR-based driving perception system is crucial for autonomous driving, as it provides real-time, accurate, and robust perception results that significantly enhance driving safety. This paper proposes a LiDAR-based panoptic driving perception algorithm named LiDAR-PDP, which can simultaneously perform free space segmentation, static object semantic segmentation, and moving object instance segmentation. This paper constructs an efficient two-branch 2D CNN backbone that consists of a spatial feature branch and a motion feature branch to extract fine-grained spatial-temporal features from sequential point cloud data. For the spatial feature branch, we designed a Compound Sampling Convolution Module (CSCM) that can learn 3D spatial features with a more general geometric modeling capacity. Additionally, This paper designs Parallel Feature Pyramid Networks (Parallel FPNs) that can aggregate multi-scale features in both range-dependent and size-dependent manners. We evaluate LiDAR-PDP on SemanticKITTI and NuScenes. Our method outperforms other comparison methods that were modified from state-of-the-art point cloud panoptic segmentation methods on both datasets, thereby demonstrating the effectiveness of the proposed components in LiDAR-PDP. Moreover, our lightweight LiDAR-PDP achieves outstanding accuracy and real-time performance on a Vehicular and Edge Computing platform. The LiDAR-PDP would improve the perception ability of autonomous vehicles with limited computing resources.