一种用于道路交通事故自动检测的改进双流网络  

Improved Two-stream Network for Vision-based Traffic Accident Detection

作　　者：王晨[1,2] 周威 严隽逸 龚耀辉 WANG Chen;ZHOU Wei;YAN Jun-yi;GONG Yao-hui(School of Transportation,Southeast University,Nanjing 211189,Jiangsu,China;Purple Mountain Laboratories,Nanjing 211111,Jiangsu,China;Qidong Municipal Transportation Bureau,Nantong 226200,Jiangsu,China)

机构地区：[1]东南大学交通学院,江苏南京211189 [2]紫金山实验室,江苏南京211111 [3]启东市交通运输局,江苏南通226200

出　　处：《中国公路学报》2023年第5期185-196,共12页China Journal of Highway and Transport

基　　金：科技部重点研发计划国际合作项目(2018YFE0102700);国家自然科学基金项目(71971061)。

摘　　要：交通事故实时快速检测对提升道路交通应急管理水平具有重要意义。视频检测技术可以有效节省人工监控成本,并在事故检测的相关研究中取得了较好成果。然而,在拥堵交通环境中,现有方法通常无法准确识别出交通事故。为解决上述问题,提出一种改进双流网络,以提高拥堵交通环境下的检测性能。该改进双流网络利用2个并行卷积神经网络(即外观、运动特征提取网络)分别提取视频数据的事故外观特征和运动特征。为了引导网络捕捉拥堵场景下判别性事故外观特征,引入辅助网络和三元损失函数来训练外观特征提取网络。在运动特征提取网络中建立光流提取网络,以捕捉拥堵交通环境下细微的光流信息,进而提升事故运动特征提取能力。研究结果表明:相较传统方法,所提方法有效提升了拥堵交通环境下的事故检测能力,降低了24.56%的误报率和25.00%的漏检率。研究成果可用于城市高流量、混流交通场景(如城市交叉口、快速路、主干道等)中进行道路交通事故自动检测。Real-time traffic accident detection is crucial to improve traffic emergency management.Vision-based methods can effectively reduce the manual monitoring costs and achieve promising results.However,existing methods typically fail to identify traffic accidents in congested traffic.To this end,this study proposes an improved two-stream network to boost detection performance in congested traffic.Specifically,the improved two-stream network utilizes two parallel convolutional neural networks(appearance and motion)to extract the appearance and motion features of traffic accidents.To guide the appearance network in capturing the prominent crash appearance features in congested traffic,we developed an auxiliary network and triplet loss to train the appearance network.Moreover,this study establishes an optical flow extraction network in a motion feature extraction network to capture subtle optical flow information.Experimental results show that the proposed method effectively improves accident detection in congested traffic,reducing the false and miss identification rates by 24.56%and 25.00%,respectively,compared to traditional methods.This method is applicable to high-volume mixed urban traffic scenarios(for example,urban intersections,expressways,and main roads).

关 键 词：交通工程 视频事故检测 改进双流网络 辅助网络 光流提取网络 

分 类 号：U491.11[交通运输工程—交通运输规划与管理]