基于深度学习的多角度车辆动态检测方法  

A Multi-aspect Method for Vehicle Dynamic Detection Based On Deep Learning

作　　者：李浩[1,2] 张运胜[2] 连捷 李泽萍 

机构地区：[1]西安文理学院西安市物联网应用工程重点实验室,西安710065 [2]东南大学交通学院,南京210000 [3]中国电子科技集团公司第38研究所,合肥230088

出　　处：《交通信息与安全》2017年第5期37-44,共8页Journal of Transport Information and Safety

基　　金：国家自然基金项目(71563045);西安市科技计划项目(CXY1531WL25)资助

摘　　要：针对在复杂场景下,背景区域干扰特征过多、被检测目标运动速度快等导致的动态目标检测率低的问题,研究了基于深度学习的多角度车辆动态检测方法,将带有微型神经网络的卷积神经网络(MLP-CNN)用于传统算法的改进。使用快速候选区域提取算法提取图像中可能存在车辆的区域,之后使用深层卷积神经网络(CNN)提取候选区域的特征,并在卷积层中增加微型神经网络(MLP)对每层的特征进一步综合抽象,最后使用支持向量机(SVM)区分目标和背景的CNN特征。实验表明,该方法能够处理高复杂度背景条件下,部分遮挡、运动速度快的目标特征检测,识别率高达87.9%,耗时仅需225ms,比常用方法效率有大幅度提升。In order to address the problems of dynamic target detection rate is low due to excessive interference of background areas and fast moving speed of detected targets in complex scenes,this article proposes a multi-aspect method for vehicle dynamic detection based on deep learning.The traditional deep learning algorithm is improved by using convolutional neural network with a multiplayer perceptron(MLP-CNN).The kernel of this improved method is first to apply the fast candidate region extraction algorithm to find the regions where vehicles may exist,then to utilize a deep convolutional neural network(CNN)to extract features of candidate region,and to use an enhanced convolutional layer with multilayer perceptron(MLP)to further abstract optimal features for each layer.The Support vector machine(SVM)is finally used to classify CNN features of backgrounds.The results show that the proposed method can deal with part occlusion or fast motion objects.With a recognition accuracy of 87.9% and elapsed time of 225 ms,it is more efficient than other traditional methods.

关 键 词：智能交通 车辆检测 深度学习 卷积神经网 微型神经网 

分 类 号：U495[交通运输工程—交通运输规划与管理]