摘要
不同于传统车辆行为识别方法大多基于鸟瞰视角下的历史轨迹信息,该文提出了一种用于自动驾驶汽车的车载视角下基于视觉信息的前车行为识别方法。针对缺乏车辆行为数据集的问题,提出了基于车载视频信息的车辆行为标注方法,构建了车载视角下的行为识别数据集。设计了一种基于SlowFast网络为主体的车辆行为识别算法,设计了焦点损失函数,并引入非局部操作模块来替换原有的交叉熵损失函数。结果表明:相较于原SlowFast模型,新模型的总体准确率提高了20.56%,实现了对视频中前方多台车辆的行为识别。
Unlike traditional vehicle behavior recognition methods,which are mostly based on historical track information from a bird's eye view,this paper proposed a new approach for autonomous vehicle behavior recognition based on visual information.A vehicle behavior labeling method based on vehicle-mounted video information was proposed,and a behavior recognition data set was constructed from vehicle-mounted perspective.A vehicle behavior recognition algorithm based on SlowFast network was designed;A Focal Loss function was designed;And a Non-local operation module was introduced.The results show that compared with the original SlowFast model,the overall accuracy of the new model is improved by 20.56%,and the behavior recognition of multiple vehicles in front of the video is realized.
作者
刘延伟
黄志明
高博麟
钟薇
陈嘉星
刘家熙
LIU Yanwei;HUANG Zhiming;GAO Bolin;ZHONG Wei;CHEN Jiaxing;LIU Jiaxi(School of Electro-Mechanical Engineering,Guangdong University of Technology,Guangzhou 510006,China;State Key Laboratory of Automotive Safety and Energy,Tsinghua University,Beijing 100084,China;National Innovation Center of Intelligent and Connected Vehicles,Beijing,100084,China)
出处
《汽车安全与节能学报》
CAS
CSCD
北大核心
2023年第6期707-714,共8页
Journal of Automotive Safety and Energy
基金
国家自然科学基金(52172389)
广东自然科学基金(2022A1515012080)。
关键词
自动驾驶汽车
车端感知
前车行为识别
SlowFast网络
车载视角
autonomous vehicles
on-board perception
front vehicle behavior recognition
SlowFast networks
vehicle perspective