V2X Perception Algorithms for Intelligent Driving: A Survey

With recent great advancements of deep learning, the perception ability of single-vehicle has obtained significant improvement. Especially, the raid developments of vision transformer, vision-centric Bird’s-Eye View (BEV) perception, and sensor fusion perception techniques have led to great progress in intelligent driving. However, the perception method applied to single-vehicle also suffers from occlusion and sparse sensor observation at a far distance due to limited perspective of environment from single-vehicle sight-of-view. To address these issues, some Vehicle-to-Everything (V2X) cooperative perception methods are proposed to leverage the advantages of multiple viewpoints of the same scene from other vehicles or infrastructure recently. These methods are superior to single-vehicle perception for intelligent driving. However, there has been no survey involving this burgeoning research field. To facilitate future research, this paper presents a comprehensive survey of developments in V2X perception methods, encompassing all the main concerns, including viewpoint, data sets, performance metrics, and the recent V2X perception methods with their strength and weakness. Further, quantitative comparisons would be provided among these methods with some in-depth analysis on still existing issues of these methods. Finally, the paper will share some insights into the issues to spark new research directions based on achievements and trends of existing methods.