AI-Empowered Performance Optimization of Cooperative Perception for Connected and Automated Vehicles

To overcome the limitations of the single-vehicle perception, cooperative perception has been proposed as a promising approach to enable the spatially separated connected and automated vehicles (CAVs) and roadside infrastructures to share and make full use of the perceptual information. However, the current works tend to perform the redundant feature interaction, neglect the spatio-temporal modeling, and lack the incentive-aware CAV selection. To overcome the drawbacks, we study the AI-empowered performance optimization of cooperative perception to achieve the high-accuracy and computation-efficient feature fusion, and utility-aware CAV selection. A novel cooperative perception framework based on the enhanced spatiotemporal interactive and sparsity-aware cooperative perception is elaborately designed to tackle the critical challenges of communication overhead, computational efficiency, and dynamic scene understanding. In addition, we present a greedy-based CAV selection approach to tackle the performance-cost tradeoff in the cooperative perception service, i.e., the overall perception performance and total incentive cost. Experimental results on the OPV2V and V2XSet datasets are presented to verify the advantages of our scheme in comparison with the benchmark schemes. Finally, we discuss the open challenges and future research directions of cooperative perception.