Cooperative Dense-Beam LiDAR-Camera Fusion Algorithm for 3D Roadside Perception

We propose an innovative fusion algorithm that combines a dense-beam LiDAR and a camera using a post-fusion strategy. This approach utilizes detection results from both images and point clouds and performs target matching based on the uniqueness of each target. Unlike traditional feature-layer fusion networks, our approach is trained using established single-sensor detection algorithms, eliminating the need for designing a feature-layer fusion network. Through experimental validation, our method surpasses the performance improvement achieved by feature-level fusion methods. Specifically, our approach achieves a significant 4.28% increase in sensing accuracy and a remarkable 25.48% improvement in pedestrian recall compared to the single-sensor system. Moreover, our method consumes fewer computational resources and is easy to implement, thereby enhancing efficiency.