Practical Collaborative Perception: Design, Implementation, and Evaluation

Multi-agent collaborative perception (CP) has recently garnered significant attention due to its potential in extending perception range and enhancing perception robustness. However, most existing CP methods have only been validated in simulations, and as far as we know, there is currently no practical CP system available. To address this gap, this paper focuses on the design, implementation, and evaluation of CP in real-world scenarios. In order to conserve communication resources, our approach advocates sharing partial point clouds rather than complete data. Furthermore, to improve computational efficiency, we propose a voxelization strategy to compute the blind zone and the required point clouds. Experiments conducted on several small autonomous vehicles demonstrate the efficiency of communication and computation.