【作者】 肖荣春； 刘元盛； 张军； 于康鸿； 张婷； 王子钰；

【Author】 XIAO Rongchun;LIU Yuansheng;ZHANG Jun;YU Kanghong;ZHANG Ting;WANG Ziyu;Beijing Key Laboratory of Information Service Engineering;College of Robotics,Beijing Union University;College of Urban and Rail Transit,Beijing Union University;

【机构】 北京市信息服务工程重点实验室； 北京联合大学机器人学院； 北京市联合大学城市与轨道交通学院；

【摘要】 从鸟瞰图(Bird’s Eye View,BEV)中学习强表示特征来完成感知任务正在成为工业界和学术界新的研究热点。自动驾驶算法的传统感知方法是在前视图或透视图中执行检测、分割、跟踪等。随着车辆上搭载的传感器变得越来越复杂,整合来自不同传感器的多源信息并在统一视图中表示特征变得至关重要。基于BEV融合感知相较于传感器后融合感知更有利于保留语义特征信息和3D空间信息,且在数据集上取得了更高的精度与更低的时延。本文回顾了有关BEV感知的最新研究成果,介绍了几种业界主流的BEV融合感知方法设计,阐明了BEV感知未来的发展方向。更多还原

【Abstract】 Learning strong representation features from bird’s eye view(BEV) to complete perceptual tasks is becoming a new research focus in industry and academia.The traditional perception method of automatic driving algorithm is to perform detection,segmentation,tracking and so on in front view or perspective view.As sensors onboard vehicles become more complex,integrating multiple sources of information from different sensors and representing features in a unified view becomes critical.BEV-based fusion perception is more conducive to preserve semantic feature information and 3 D spatial information than post-sensor fusion perception,and achieves higher accuracy and lower time delay on data sets.This paper reviews the latest research results of BEV perception,introduces several mainstream BEV fusion sensing methods,and expounds the future development direction of BEV perception.更多还原