基于卷积神经网络的运动行人识别算法的研究  

Research on moving pedestrian recognition algorithm based on convolutional neural network

作　　者：于闯 杨姝[2] 寇海莲[2] YU Chuang;YANG Shu;KOU Haitian(Computer and Basic Mathematics Education Department,Shenyang Normal University,Shenyang 110034,China;College of Educational Technology,Shenyang Normal University,Shenyang 110034,China)

机构地区：[1]沈阳师范大学计算机与数学基础教学部,沈阳110034 [2]沈阳师范大学教育技术学院,沈阳110034

出　　处：《沈阳师范大学学报（自然科学版）》2019年第5期461-466,共6页Journal of Shenyang Normal University:Natural Science Edition

基　　金：全国教育科学“十三五”规划2019年度教育部重点项目(DCA190329);辽宁省教育厅人文社会科学研究项目(W201613)

摘　　要：针对定点视频监控中,因位于远景区域中的行人目标分辨率较低等因素容易造成行人漏检和误检的问题,提出一种对监控系统中的远景区域增设辅助变焦摄像头的方法,来解决此问题。具体的方法如下:首先,增设辅助变焦摄像头对系统中的远景区域进行监控,使得位于远景区域中的视频图像分辨率提高至原来的2倍;其次,对监控系统中远景区域和增加的辅助摄像头获得的视频图像分别进行运动目标的检测、去噪和提取,得到运动目标的图像;然后,通过人工方法,对图像中的运动目标进行类别标识,获得训练样本和检验样本;最后,利用训练样本训练卷积神经网络,进行运动行人的识别。实验结果表明,使用本文提出的方法,可以有效提高远景区域中运动行人的识别准确率,卷积神经网络准确率达到92.61%,行人目标的相对检出率提高了95.29%,行人分类准确率提高了15.14%,可以满足识别精度的要求。For fixed-point video surveillance systems, due to the low resolution of pedestrian targets located in the far scene area and other factors, missed detection and false detection during pedestrian detection process occur frequently. To solve this problem, a method for adding an auxiliary zoom camera to the far scene area of the monitoring system is proposed. The specific method is as follows: Firstly, an auxiliary zoom camera is added to monitor the far scene area in the system, so that the resolution of the video images located in the far scene area is increased by 2 times;secondly, after the auxiliary camera is added to monitor the far scene area of the monitoring system, the obtained video images are respectively detected, denoised and extracted;then, the moving targets detected in the video images are identified by a manual method to obtain training samples and testing samples;finally, the training samples are utilized to train a convolutional neural network to identify pedestrians. The experimental results show that the proposed method can effectively improve the recognition accuracy of pedestrians in the far scene area. The accuracy of convolutional neural network is 92.61%, the relative detection rate of pedestrian targets is increased by 95.29%, and the accuracy of pedestrian classification is increased by 15.14%, it can meet the requirements of recognition accuracy.

关 键 词：深度学习 卷积神经网络 行人检测 

分 类 号：TP319[自动化与计算机技术—计算机软件与理论]