V2IViewer: Towards Efficient Collaborative Perception via Point Cloud Data Fusion and Vehicle-to-Infrastructure Communications

Collaborative perception (CP) with vehicle-to-infrastructure (V2I) communications is a critical scenario in high-level autonomous driving. This paper presents a novel framework called V2IViewer to facilitate collaborative perception, which consists of three modules: object detection and tracking, data transmission, and object alignment. On this basis, we design a heterogeneous multi-agent middle layer (HMML) as the backbone to extract feature representations, and utilize a Kalman filter (KF) with the Hungarian algorithm for object tracking. For transmitting object information from infrastructure to ego-vehicle, Protobuf is utilized for data serialization using binary encoding, which reduces communication overheads. For object alignment from multiple agents, a Spatiotemporal Asynchronous Fusion (SAF) method is proposed, which uses a Multilayer Perceptron (MLP) for generating post-synchronization object sequences. These sequences are then utilized for fusion to enhance the accuracy of the integration. Experimental validation on DAIR-V2X-C, V2X-Seq, and V2XSet datasets shows that V2IViewer enhances long-range object detection accuracy by an average of 12.9% over state-of-the-art collaborative methods. Moreover, V2IViewer demonstrates an average improvement in accuracy of 3.3% across various noise conditions compared to existing models. Finally, the system prototype is implemented and the performance has been validated in realistic environments.