Optimized Collaborative Perception: Sector-Based BEV Fusion in Limited Communication Conditions

Collaborative perception is essential in autonomous driving, enabling connected autonomous vehicles (CAVs) to share sensor data and improve awareness of their surroundings. This is especially critical for detecting occluded objects at intersections, where limited visibility can compromise safety and hinder real-time decision making. However, early fusion of sensor data across multiple CAVs presents significant challenges in data management, as the sheer volume of 3D point cloud information demands substantial communication bandwidth. To address these challenges, this article proposes an optimized collaborative perception framework specifically designed for CAVs at intersections. Our approach begins with each CAV generating a Bird's Eye View (BEV) map from LiDAR data, which is then divided into sectors. The quality and size of the data for each sector are assessed and sent to a roadside unit (RSU) that acts as a data center. The RSU selectively coordinates the sharing of high-quality sectors only, reducing redundant data transmission by avoiding empty or low-density regions. Through this targeted data sharing, our framework minimizes communication loads and computational demands while preserving perception accuracy, thus supporting efficient and scalable collaborative perception in complex intersection environments.