DA-ROI: A Multimodal Environment Perception Approach with Depth Adaptive ROI Scaling for Safe Autonomous Driving

Accurate 3D perception is crucial for intelligent transportation systems, enabling applications from autonomous driving to vehicle cooperation. However, LiDAR-camera fusion suffers from projection uncertainty caused by calibration errors and depth noise, leading to misalignment that compromises perception reliability. This results in incomplete coverage of distant vulnerable road users and noise contamination for nearby vehicles, increasing false negatives and false positives and posing safety risks. To address this, we propose DA-ROI, a depth-aware adaptive ROI scaling method that reframes ROI adjustment as data-driven uncertainty compensation. It employs robust depth estimation and a lightweight coupling function to dynamically determine scaling ratios based on target depth and size, balancing feature coverage and noise suppression. Evaluations on KITTI and DAIR-V2X demonstrate that DA-ROI consistently improves detection accuracy across various fusion models, with significant gains in vulnerable road user recognition. The method provides an efficient solution for enhancing perception robustness in intelligent transportation systems.