基于改进DRDL模型的车辆再识别算法  

Improved Vehicle Re-Identification Algorithm Based on DRDL Model

作　　者：杨飚[1] 王雨周 YANG Biao;WANG Yu-zhou(Beijing Key Laboratory of Urban Road Traffic Intelligent Control Technology,North China University of Technology,Beijing 100144)

机构地区：[1]北方工业大学城市道路交通智能控制技术北京市重点实验室

出　　处：《现代计算机》2019年第28期34-38,共5页Modern Computer

基　　金：国家自然科学基金资助项目(No.61374191);北京市教育委员会科技计划资助项目(No.KM201710009001)

摘　　要：为了提高车辆再识别的准确率,设计并实现一种基于改进DRDL模型的车辆再识别算法:算法整体采用DRDL模型框架结构,首先改进车辆的特征提取部分,采用ResNet残差网络,将提取到的1000维特征分别进行车辆型号分类和车辆ID识别的训练;其次,根据人脸检测算法研究,提出将ARC loss应用于车辆度量学习中。实际实验表明:在Ve hicleID数据集上基于改进DRDL模型的车辆再识别算法准确度能够提升14.3%,可以达到63.2%,车辆型号识别准确度能够提升14.4%,达到97.7%,优于原DRDL模型算法。In order to improve the accuracy of vehicle re-identification,designs and implements the vehicle re-identification algorithm based on im proved DRDL(Deep Relative Distance Learning)model.This algorithm adopts DRDL model framework structure.Firstly,the vehicle fea ture extraction part is replaced by ResNet,which extracts 1000-dimensional features that are separately trained for vehicle model classifi cation and vehicle ID identification.Secondly,according to the face detection algorithm research,proposes to apply ARC loss to vehicle metric learning.The actual experiment shows that the accuracy of the vehicle re-identification algorithm based on the improved DRDL model on the VehicleID data set can be improved by 14.3%,which can reach 63.2%,and the vehicle model identification accuracy can be improved by 14.4%to 97.7%,which is better than the original DRDL model algorithm.

关 键 词：车辆再识别 卷积神经网络 车辆外观特征 细粒度局部特征 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]