Toward Robust Cooperative Perception via Spatio-Temporal Modelling

Cooperative perception as an emerging application of LiDAR-driven signal processing in driving scenarios has received widespread attention in recent years. Despite impressive advancements in previous works through sophisticated strategies, challenges remain due to inevitable data sparsity and localization errors. To this end, we propose a Spatio-Temporal Cooperative Perception (STCP) framework to address the above issues. Our novelties derive from two core components. A multi-scale temporal integration module is introduced to aggregate historical clues from the ego agent for mitigating data sparsity interference. In addition, we design a spatial cooperation transformer to perform pragmatic cooperation and eliminate the feature misalignment from collaborators due to localization errors. Extensive experiments are conducted on real-world and simulated multi-agent 3D object detection datasets. Quantitative analyses show that our framework outperforms existing methods on DAIR-V2X and V2X-Sim datasets with significant gains of 2.16% and 2.98% regarding AP@0.5.