PAFNet: Pillar Attention Fusion Network for Vehicle–Infrastructure Cooperative Target Detection Using LiDAR

With the development of autonomous driving, consensus is gradually forming around vehicle–infrastructure cooperative (VIC) autonomous driving. The VIC environment-sensing system uses roadside sensors in collaboration with automotive sensors to capture traffic target information symmetrically from both the roadside and the vehicle, thus extending the perception capabilities of autonomous driving vehicles. However, the current target detection accuracy for feature fusion based on roadside LiDAR and automotive LiDAR is relatively low, making it difficult to satisfy the sensing requirements of autonomous vehicles. This paper proposes PAFNet, a VIC pillar attention fusion network for target detection, aimed at improving LiDAR target detection accuracy under feature fusion. The proposed spatial and temporal cooperative fusion preprocessing method ensures the accuracy of the fused features through frame matching and coordinate transformation of the point cloud. In addition, this paper introduces the first anchor-free method for 3D target detection for VIC feature fusion, using a centroid-based approach for target detection. In the feature fusion stage, we propose the grid attention feature fusion method. This method uses the spatial feature attention mechanism to fuse the roadside and vehicle-side features. The experiment on the DAIR-V2X-C dataset shows that PAFNet achieved a 6.92% higher detection accuracy in 3D target detection than FFNet in urban scenes.