深度分类网络研究及其在智能视频监控系统中的应用  

Deep Classification Networks and Its Application in Intelligent Video Surveillance System

作　　者：孙宁[1] 陈梁[1] 韩光[1] 李晓飞[1] 

机构地区：[1]南京邮电大学宽带无线通信技术教育部工程研究中心,南京210003

出　　处：《电光与控制》2015年第9期77-82,共6页Electronics Optics & Control

基　　金：国家自然科学基金(61471206);江苏省自然科学基金(BK20141428);南京邮电大学引进人才项目(NY212037)

摘　　要：研究了深度分类网络在道路交通典型目标分类中的应用,使用原始灰度图、HOG特征直方图、Canny边缘图与本征特征等多种目标表征方法与深度置信网络(Deep Belief Networks,DBN)相结合构建深度分类网络实现对行人、骑车人、车辆和其他4种典型道路交通目标的分类功能。为了配合基于DBN的深度人车分类网络的训练,建立了称为NUPTERC的典型道路目标图像库,给出了建库的规则和方法,利用NUPTERC图像库构建实验对深度分类网络进行测试,并与其他典型人车分类方法进行了比较。证明深度分类网络在满足实时性的条件下,可以获得令人满意的分类正确率。最后,将基于DBN5Canny的人车分类算法应用于智能视频分析云平台,实现了对道路上的典型目标实时、精确的统计和分类功能。Application of deep classification networks in classification of typical targets in road traffic is investigated in this paper. Deep classification networks are constructed by combining such target representation methods as original gray-level image, HOG feature histogram, Canny edge image and eigen-features with Deep Belief Networks （DBN）, to realize the classification function for four typical targets in road traffic： pedestrian, biker, vehicles and others in the real scene. In order to assist in training of DBN-based deep people/ vehicle classification networks, an image database of typical road targets called NUPTERC is established, with rules and methods for its establishment. And then experiments are constructed with NUPTERC image database, to test the proposed deep classification networks, and a comparison is made with other classification methods for people and vehicles. It is proven that the deep classification networks can achieve satisfactory classification accuracy under the condition of meeting the real-time performance. Finally, people/vehicles classification 7V5 algorithm based on DB Canny is applied to the ＂cloud platform for intelligent video analysis＂ developed by our center, realizing functions of real-time accurate analysis and classification of typical targets in road traffic.

关 键 词：目标分类 深度置信网络 特征提取 智能视频监控系统 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]