改进残差网络的多角度车型识别方法  

Improve dresidual network for multi-angle vehicles recognition

作　　者：蒋行国[1] 苏欣欣 蔡晓东[1] JIANG Xingguo;SU Xinxin;CAI Xiaodong(School of Information and Communicatio,Guilin University of Electronic Technology,Guilin 551004,China)

机构地区：[1]桂林电子科技大学信息与通信学院,广西桂林541004

出　　处：《电视技术》2018年第6期93-98,共6页Video Engineering

基　　金：2016年广西物联网技术及产业化推进协同创新中心资助项目(WLW200601);2016年"认知无线电与信息处理"省部共建教育部重点实验室基金项目(CRKL160102)

摘　　要：在现实交通场景中,现有车型识别方法主要针对正面或侧面角度的车辆,但由于识别角度相对单一并不适用于多角度的车型识别。为满足实际场景下对车型识别要求,提出一种改进的残差结构特征提取网络,对其结构进行加宽改进,网络使用较少参数提取特征,加快整体网络的收敛速度。其次,结合使用基于可调类间距的Softmax Loss度量学习方法(Large-Margin Softmax Loss)进行车型识别,达到增大类间距离并减小类内距离的学习目标,提高识别的准确率。实验表明,本方法能够在交叉路口、林荫道、园区道路等复杂交通场景下进行多角度车型识别,测试识别准确率达97.4%。In real traffic scenarios, the existing recognition methods mainly aim at the vehicle with frontal or side augles, and as the recognition angle is relatively simple, they are not suitable for multi - angle recognition. For real applications, a multi - angle model recognition method is proposed. Firstly, a teature extraction network based on residual structure is proposed, it widens the network and uses less parameters to speed up the convergence of the whole network for feature extraction. Secondly, vehicle recognition combined with the tone pitch Softmax Loss metric learning method （ Large - Margin Softmax Loss） is applied to increase the distance betweent classes and reduce the distance betweent objects in a class, so as to improve the accuracy of recognition. Experimental resuhs show that this method can classify multi -angle vehicle types in complex traffic scenes such as interseetion, boulevard, park road and so on, and the recognition accuracy can reach up to 97.4%.

关 键 词：残差结构 卷积神经网络 度量学习 复杂交通场景 车型识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]