Multi-Scale Dynamic Spatial Attention Module for Robust Point Cloud Perception in Cooperative Vehicle Infrastructure System

Abstract

Cooperative Vehicle-Infrastructure Systems (CVIS) enhance perception and safety by enabling information exchange between vehicles and roadside units. LiDAR is a key sensor in CVIS due to its high accuracy and resilience to lighting conditions. However, limited computational resources present challenges in effectively detecting objects of varying sizes, achieving accurate localization, and maintaining stability in occluded or dense traffic scenarios. This paper proposes the Multi-Scale Dynamic Spatial Attention Module (MSDSAM) to address these issues. MSDSAM first divides the point cloud into multi-scale pillars to capture richer spatial features. Then, a dynamic attention mechanism adaptively fuses features across scales, improving detection robustness and efficiency. The module enhances detection performance for both small and large objects while reducing computational overhead. MSDSAM is model-agnostic and can be integrated into existing pillar-based point cloud detectors. Extensive experiments on both the DAIR-V2X-C and V2XSet datasets demonstrate that MSDSAM consistently improves detection accuracy in both single-agent and cooperative scenarios, fully showcasing its effectiveness and generalizability. The code will be released at https://github.com/usergxx/MSDSAM