CmpV2X: Enabling Communication-efficient and Robust Cooperative Perception for Internet-of-Vehicles

As a promising multi-agent perception paradigm, feature-level cooperative perception has attracted widespread attention due to its enhancement of the perception capabilities of autonomous vehicles. However, existing solutions invariably ignore the heterogeneity in computational resources and local features among diverse agents, resulting in sub-optimal collaboration performance. In this paper, we propose CmpV2X, a communication-efficient and robust multi-agent perception system designed to address resource allocation and feature fusion challenges in dynamic Internet of Vehicles (IoV) scenarios. Specifically, we first introduce a resource-guided model selection method that dynamically select perception models according to the computational resources of agents to meet latency requirements. Second, we propose a three-stage communication scheme that leverages confidence maps to select agents for collaboration, enabling efficient communication based on data importance. Finally, an inter-agent domain adaptation and feature fusion module is presented to mitigate data heterogeneity among agents and achieve robust information aggregation. With existing cooperative perception datasets, we simulate dynamic IoV scenarios where agents adopt different perception models and exhibit varying computational resources. Extensive experiments demonstrate that CmpV2X significantly improves collaborative 3D object detection performance and improves average resource utilization compared to state-of-the-art methods.