采用深度卷积神经网络技术的汽车群体属性分析研究  

Vehicle Group Attributive Analysis Based on Deep Convolution Neural Network Technology

作　　者：任军利 冉崇善[2] REN Jun-li;RAN Chong-shan(Information and Network Management Center,Shaanxi University of Science ＆ Technology,Xi＇an 710021;School of Electrical and Information Engineering,Shaanxi University of Science ＆ Technology,Xi＇an 710021 China)

机构地区：[1]陕西科技大学信息与网络管理中心,陕西西安710021 [2]陕西科技大学电气与信息工程学院,陕西西安710021

出　　处：《湘潭大学学报（自然科学版）》2018年第4期91-96,共6页Journal of Xiangtan University(Natural Science Edition)

基　　金：高校资源流量优化系统的服务支撑与应用测试项目(210160097)

摘　　要：针对交通监控视频中的车道线相互不平行且无法完整贴合车道导致车流量无法准确统计这一问题,提出了一种以深度卷积神经网络技术为框架进行车流量统计的算法.算法基于虚拟线圈和DCNN,通过前景分析、连通域分析和基于DCNN的车辆跨线行驶判决,大幅度提高了车流量统计的准确率,通过实际监控视频中的随机测试进行了算法实验,实验结果表明基于虚拟线圈和DCNN的车流量统计算法的相对准确度均值达到了98.05%,绝对准确度均值达到了88.5%.Aiming at the problem that the traffic flow cannot be accurately counted because the lanes in traffic surveillance video are not parallel to each other and cannot fully fit the lanes, an algorithm based on deep convolution neural network technology is proposed. The algorithm is based on virtual coil and DCNN. The accuracy of traffic flow statistics is greatly improved by foreground analysis, connected area analysis and DCNN-based vehicle crossing decision. The algorithm is tested by random test in real monitoring video. The experimental results show that the phase of traffic flow statistics algorithm based on virtual coil and DCNN reached the accuracy of 98.05%, and the absolute accuracy reached 88.5%.

关 键 词：汽车群体属性 车流量统计 深度卷积神经网络 虚拟线圈 

分 类 号：TG202[金属学及工艺—铸造]