Optimizing roadside LiDAR beam distribution to enhance vehicle detection performance considering dynamic vehicle occlusion effects

The distribution of LiDAR beams is the primary factor determining the density and coverage of point clouds over target vehicles, thus it has the most significant impact on vehicle detection performance. However, quantifying the influence of beam configurations on perception outcomes in complex traffic environments is challenging due to dynamic occlusions between vehicles and varying traffic densities. This study provides an effective framework for optimizing LiDAR beam distribution in dynamic traffic environments. Firstly, we proposed a dynamic occlusion model to calculate expected occlusion effects of target vehicles under various traffic flow densities. Based on this, an analytical model is developed to quantify the relationship between LiDAR beam distribution and vehicle detection performance. Furthermore, we developed a highly efficient optimization model for LiDAR beam distribution to enhance vehicle detection performance. Real-world experiments were conducted to collect vehicle point clouds using two types of LiDAR across six scenarios to validate the proposed models. Additionally, a series of simulation-based experiments demonstrated that the LiDAR beam distributions obtained from the optimization model achieved superior vehicle detection performance compared to SOTA methods. Notably, the optimization results for the 16-beam LiDAR are particularly significant, enhancing vehicle detection Recall by up to 63.7% (a 4.4-fold increase) compared to the baseline beam distribution. In addition, the optimization for an 80-beam LiDAR is completed in just 45 s, representing a speed improvement of 1,000 times compared to SOTA methods. This work provides both theoretical and practical contributions to the design of LiDAR sensing systems, which can be applied to intelligent transportation systems and autonomous driving infrastructure.