摘要
高速公路异常事件自动检测是有效保障道路交通安全和运输效率的重要手段,由于监控视频数据量巨大,现有自动检测算法存在实时性、准确性低的问题。为此本文提出了基于轨迹分类的对比性悲观似然(comparative pessimistic likelihood estimation,CPLE)算法。构建了包含车辆检测、车辆跟踪和轨迹分类3种功能的异常事件自动检测模型框架,采用YOLO v3对车辆进行目标检测,获得4类不同车辆类型的相关信息,采用简单在线和实时跟踪算法对车辆进行多目标跟踪,获得不同场景的异常事件车辆轨迹;基于半监督学习,采用极大似然法对车辆轨迹分类进行改进,引入对比性悲观似然估计,围绕其对比和悲观原则进行参数设置和标定,进行异常事件轨迹分类和确认,提出基于车辆轨迹的异常事件自动检测算法。以甘肃省G312线公路智能化检测系统为测试对象,共收集1300段视频,形成530条测试集轨迹和630条验证集轨迹,测试结果表明:通过对不同场景异常事件进行检测和预警,基于对比性悲观似然估计的轨迹分类算法性能准确率达到89.7%,比自学习和监督学习方法的准确率分别高出23.6%和41.3%,尽管对散落货物和超速事件的检测正确性稍低,平均为77.0%,但突发性停车、拥堵和事故的检测平均正确率达98.2%,在严重影响交通的事件检测方面的平均正确率达到94%。本方法丰富了高速公路异常事件自动检测算法,可作为异常事件自动检测提供备选方法。
An automatic freeway incident detection method is important for maintaining a safe,efficient traffic operation.Due to the fact that a large number of surveillance videos may hinder the real-time and accurate response of current automatic incident detection algorithms,a comparative pessimistic likelihood estimation(CPLE)algorithm based on trajectory classification is proposed.A framework for automatic detection of anomalous events,which contains vehicle detection,vehicle tracking and trajectory classification,is developed.YOLO v3 is employed to detect the vehicles,and related information about four different types of vehicles is obtained.Online real-time tracking algorithms are used for multi-target tracking of vehicles.Anomalous event vehicle trajectories are obtained for different scenarios.Based on semi-supervised learning,the maximum likelihood method is employed to improve the classification of vehicle trajectories.CPLE is introduced and parameter setting and labeling are centered on comparison and pessimistic rules in order to classify and determine the incident trajectories,consequently,the automatic incident detection algorithm based on vehicle trajectories is proposed.The intelligent inspection system of Gansu Province G312 highway is used as a test object.A total of 1300 videos were collected.Among them,530 and 630 tracks are employed as test set and validation set,respectively.By testing difference scenarios of incidents and prewarning,the algorithm accuracy of trajectory classification based on CPLE reaches 89.7%,which is 23.6%higher than that of self-learning and 41.3%higher than that of supervised learning,respectively.Although the accuracy of scattered goods and speeding is averaged about 77.0%,the accuracy of sudden stopping,congestion,and accidents reaches 98.2%,and as for the incident detection influencing traffic seriously,the average accuracy reaches 94%.The proposed method enriches automatic incident detection algorithms and can be considered an alternative for freeway incident detection.
作者
李斌
马静
徐学才
马昌喜
LI Bin;MA Jing;XU Xuecai;MA Changxi(Lanzhou LongKing Transportation Science&Technology Co.Ltd.,Lanzhou 730030,China;School of Civil and Hydraulic Engineering,Huazhong University of Science and Technology,Wuhan 430074,China;School of Traffic and Transportation Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)
出处
《交通信息与安全》
CSCD
北大核心
2023年第3期23-29,共7页
Journal of Transport Information and Safety
基金
国家自然科学基金项目(52062027、72131008)
甘肃省科技重大专项计划项目(22ZD6GA010)资助。
