Center-Aware 3D Object Detection with Attention Mechanism Based on Roadside LiDAR

Infrastructure 3D Object Detection is a pivotal component of Vehicle-Infrastructure Cooperated Autonomous Driving (VICAD). As turning objects account for a high proportion of traffic at intersections, anchor-free representation in the bird's-eye view (BEV) is more suitable for roadside 3D detection. In this work, we propose CetrRoad, a simple yet effective center-aware detector with transformer-based detection head for roadside 3D object detection with single LiDAR (Light Detection and Ranging). CetrRoad firstly utilizes a voxel-based roadside LiDAR feature encoder module that voxelizes and projects the raw point cloud into BEV with dense feature representation, following a one-stage center proposal module that initializes center candidates of objects based on the top N points in the BEV target heatmap with unnormalized 2D Gaussian. Then, taking attending center proposals as query embedding, a detection head with multi-head self-attention and multi-scale multi-head deformable cross attention can refine and predict 3D bounding boxes for different classes moving/parked at the intersection. Extensive experiments and analyses demonstrate that our method achieves state-of-the-art performance on the DAIR-V2X-I benchmark with an acceptable training time cost, especially for Car and Cyclist. CetrRoad also reaches comparable results with the multi-modal fusion method for Pedestrian. An ablation study demonstrates that center-aware query as input can provide denser supervision than a purified feature map in the attention-based detection head. Moreover, we were able to intuitively observe that in complex traffic environment, our proposed model could produce more accurate 3D detection results than other compared methods with fewer false positives, which is helpful for other downstream VICAD tasks.