BEV perception for autonomous driving: State of the art and future perspectives

Abstract

The remarkable performance of Bird’s Eye View (BEV) in perception tasks has led to its gradual emergence as a focal point of attention in both industry and academia. Environmental information perception technology represents a core challenge in the field of autonomous driving, and traditional autonomous driving perception algorithms typically perform tasks such as detection, segmentation, and tracking from a frontal or specific viewpoint. As the complexity of sensor parameters configured on vehicles increases, it has become crucial to integrate multi-source information from different sensors and present features in a unified view. BEV perception is favored because it is an intuitive and user-friendly way to fuse information about the surrounding environment and provide an ideal object representation for subsequent planning and control modules. However, BEV perception also faces some key challenges. One such challenge is how to convert from a perspective view to a BEV view while reconstructing lost 3D information. The question of how to obtain accurate ground truth annotations in the BEV grid is of great importance. Similarly, the design of effective methods to integrate features from different sources is a crucial aspect of BEV perception. In this paper, we first discuss the inherent advantages of BEV perception and introduce the mainstream datasets and performance evaluation criteria for BEV perception. Furthermore, we present a comprehensive examination of recent research on BEV perception from four distinct perspectives, exploring a range of solutions, including BEV camera, BEV LiDAR, BEV fusion, and V2V multi-vehicle cooperative BEV perception. Finally, we identify prospective research directions and challenges in this field, with the aim of providing inspiration to related researchers.