基于改进ViBe和机器学习的行人头肩检测方法

发布时间：2018-04-25 02:35

本文选题：联动标定 + ViBe　；参考：《华东交通大学》2017年硕士论文

【摘要】：现如今,视频监控技术广泛应用于银行、超市、车站以及学校等公共区域的行人监测。但是现有视频监控系统仅处于对视频的记录阶段,无法精准捕捉运动行人及其清晰外貌,这给公安刑侦人员调查取证带来一定困难。因此,研究视频监控中的行人检测与清晰外貌捕捉方法具有重大意义。头肩检测作为行人外貌捕捉的关键步骤,其目的在于准确地获取行人头肩位置,为行人清晰外貌的捕捉提供可靠的前提条件。本文应用图像处理与机器学习技术,开展了公共区域下行人头肩检测方法研究,主要研究内容分为三个模块,分别是运动目标检测模块、行人头肩检测模块以及主从摄像机联动标定模块,具体工作及创新点如下:(1)针对行人清晰外貌捕捉要求,采用鱼眼摄像机结合PTZ摄像机设计主从式监控系统,同时针对主从摄像机之间的联动要求,应用数据拟合的空间标定算法,在合理选取样本点的基础上,依据样本点在鱼眼摄像机中的像素位置和PTZ摄像机拍摄样本点所需旋转的角度生成查找表,完成主从摄像机之间的联动标定,实现了主从摄像机之间的联动要求。在整体行人头肩检测与清晰外貌捕捉系统搭建完成后的验证表明,每5ms就可以完成一次主从摄像机之间的联动。(2)针对传统ViBe算法中存在的“死区”以及运动目标阴影干扰问题,提出了结合感知哈希算法和基于图像RGB色彩信息的高斯拉普拉斯差分算法的改进ViBe算法,实现了对“死区”的抑制,消除了运动目标阴影,完成了对视场范围内运动目标的检测。实验表明,相较于传统ViBe算法在视频的1315帧才能完成对“死区”的抑制,改进后的ViBe算法仅在视频的15帧就完成了对“死区”的抑制,同时没有了运动阴影的干扰。(3)针对行人头肩检测过程中误检率高的问题,提出两阶段头肩检测算法:首先,使用基于AdaBoost思想的级联分类算法训练HOG特征生成第一阶段头肩检测器,检测出行人头肩部位的“候选区域”;接着,使用SVM分类算法训练ORB特征生成第二阶段头肩检测器,对“候选区域”进行第二次检测,并以此为最终结果。实验表明,两阶段检测算法头肩检测准确率达到了80.86%,相较于传统HOG+AdaBoost检测算法准确率提升了近10个百分点。
[Abstract]:Today, video surveillance technology is widely used in banks, supermarkets, stations, schools and other public areas of pedestrian monitoring. However, the existing video surveillance system is only in the recording stage of the video, which can not accurately capture the movement of pedestrians and their clear appearance, which makes it difficult for police investigators to investigate and obtain evidence. Therefore, it is of great significance to study pedestrian detection and clear appearance capture in video surveillance. Head-shoulder detection is a key step in pedestrian appearance capture, which aims at accurately obtaining the position of head and shoulder of pedestrians, and provides a reliable precondition for the capture of clear appearance of pedestrians. In this paper, using image processing and machine learning technology, the research of human head and shoulder detection method in public area downlink is carried out. The main research content is divided into three modules, which are moving target detection module. Pedestrian head-shoulder detection module and master-slave camera linkage calibration module. The specific work and innovation are as follows: 1) aiming at the requirements of pedestrian clear appearance capture, a master-slave monitoring system is designed with a fish-eye camera combined with a PTZ camera. At the same time, according to the requirements of the linkage between the master and slave cameras, the spatial calibration algorithm of data fitting is applied to select the sample points reasonably. According to the pixel position of the sample point in the fish-eye camera and the angle of rotation needed by the PTZ camera to shoot the sample point, the look-up table is generated, and the linkage calibration between the master-slave camera and the master-slave camera is completed, and the linkage requirement between the master-slave camera is realized. The verification of the whole pedestrian head-shoulder detection system and the clear appearance capture system shows that every 5ms can complete the linkage between the master and slave cameras once. (2) aiming at the problem of "dead zone" and shadow interference of moving targets in the traditional ViBe algorithm, An improved ViBe algorithm based on perceptual hashing algorithm and Gao Si Laplace difference algorithm based on image RGB color information is proposed, which can suppress the dead zone and eliminate the shadow of moving target. The detection of moving targets in the field of view is completed. The experimental results show that compared with the traditional ViBe algorithm, the "dead zone" can be suppressed only in 15 frames of the video by the improved ViBe algorithm, which can suppress the "dead zone" only in the 1315 frames of the video, and the improved ViBe algorithm can suppress the "dead zone" only in 15 frames of the video. At the same time, there is no interference of moving shadow. 3) aiming at the problem of high false detection rate in pedestrian head-shoulder detection, a two-stage head-shoulder detection algorithm is proposed: first, The concatenated classification algorithm based on AdaBoost is used to train the HOG feature to generate the first stage head-shoulder detector to detect the "candidate area" in the head and shoulder position of the traveller, and then the SVM classification algorithm is used to train the ORB feature to generate the second stage head-shoulder detector. The candidate regions were tested for a second time, and the results were taken as the final results. The experimental results show that the accuracy of head and shoulder detection is 80.86%, which is 10 percentage points higher than that of traditional HOG AdaBoost detection algorithm.
【学位授予单位】：华东交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP181

【参考文献】