Region-Based Hybrid Collaborative Perception for Connected Autonomous Vehicles

Collaborative perception is considered as an effective approach in solving the problem for the limited fields-of-view of single vehicle. Recent studies focused on single collaboration strategy (i.e., early, intermediate, or late collaboration). However, it is difficult to achieve a balance between accuracy and transmission data size. To overcome this limitation, a region-based hybrid collaborative perception architecture for connected autonomous vehicles is proposed in this study. Specifically, data is divided into two types according to the overlapping area between the detection ranges of vehicles. For the overlapping area, intermediate collaboration is applied by sharing and fusing the features from different vehicles. The fusion model comprises the proposed cross-agents attention and global context attention modules to adaptively highlight the critical features and capture the long-range dependency, thereby improving the detection accuracy. Moreover, an object detection head based on the Gaussian mixture model (GMM) is proposed to enhance the robustness to location noise. For the non-overlapping area, late collaboration is conducted by generating and sharing the local detection result with an economic bandwidth. Finally, the experiment with the OPV2V dataset is conducted to evaluate the performance of the proposed architecture. The results show that the proposed architecture achieved 83.6% AP, outperforming the state-of-art methods. Additionally, the experiments also prove that the network with GMM module is more robust to noise.