Accelerating Collaborative Perception via Cooperative Inference in Vehicular Edge Computing

Recent years have witnessed significant advancements in collaborative perception, particularly in enhancing accuracy and reducing communication overhead. However, due to the limited computation capacity of vehicles and dynamic bandwidth environment, traditional methods are constrained by fixed computational strategies, posing a challenge to providing low delay services in heterogeneous Vehicular Edge Computing (VEC). Considering this, we present a framework to accelerate collaborative perception, where vehicles adaptively partition the inference models and jointly offload them to the edge node. Further, we model the total delay by considering task arrival order, as well as the heterogeneous capacities of each node. Then, we formulate the Model Partitioning and Offloading (MPO) problem aiming to minimize the total delay of collaborative perception tasks. On this basis, we propose the Partitioning and Offloading Points Selection (POPS) algorithm, leveraging dynamic Thompson Sampling to select the optimal offloading points for each vehicle. By actively adjusting exploration intensity and passive parameter update rules, the POPS algorithm is highly adaptable to dynamic bandwidth environments. Finally, we conduct extensive performance evaluations, and the results demonstrate the superiority of our algorithm.