Collaborative perception has emerged as a key design consideration for autonomous driving vehicles, which may suffer from two essential challenges: overwhelming communication overhead due to the sharing of sensing data, and stringent delay constraints imposed by highly dynamic traffic scenarios. To address these challenges, this paper first proposes a multi-vehicle collaborative perception framework for onboard sensing enhancement via vehicle-to-vehicle (V2V) communication aiming at reducing communication overhead and delay by selecting appropriate vehicles for collaboration. Then, the collaborative delay and the collaborative value are modeled by jointly considering the number of sensing objects and the size of sensing area. On this basis, we formulate an optimal collaborator selection (OCS) problem, which aims to maximize the collaborative value while satisfying delay constraints. Further, we design an adaptive selection algorithm (ASA), which first groups the assisting vehicles by considering their locations to reduce the overlap between different groups. Then, the vehicle selection is determined based on multi-armed bandit (MAB) learning. Finally, we conduct both simulation and field experiments, which demonstrate that ASA can improve collaborative efficiency by about 30.69% in HighD and about 32.30% in InD compared with existing competitive methods.