V2X-MoST: Motion-Aware Spatio-Temporal Compensation for Robust V2X Cooperative 3D Object Detection

Vehicle-to-everything (V2X) cooperative perception improves autonomous-driving perception by sharing complementary observations from connected vehicles and roadside infrastructure. However, communication latency causes motion-dependent feature misalignment, whereas temporary occlusion and sparse LiDAR observations produce weak or missing bird’s-eye-view (BEV) responses. This paper proposes V2X-MoST, a closed-loop motion-aware spatio-temporal compensation framework for robust LiDAR-based cooperative 3D detection. Its Motion-Aware Feature Compensation (MAFC) module aligns delayed cooperative features by combining a physics-inspired displacement derived from the previous BEV velocity field with a learnable residual offset inferred from sparse feature correlations. An Occlusion-Aware Temporal Dynamics Enhancement Head (OA-TDEH) integrates temporal memory, motion priors, and occlusion-aware gating to strengthen weak responses. The estimated velocity field is fed back to MAFC at the next frame, forming a closed-loop process. Experiments on V2XSet and DAIR-V2X validate the framework in simulated and real-world scenarios. Under the V2XSet Noisy setting, V2X-MoST achieves 0.909 [email protected] and 0.701 [email protected] and remains stable as delay increases. On DAIR-V2X, it reaches 0.670 [email protected] and 0.510 [email protected], demonstrating robust performance under vehicle-infrastructure cooperation.