基于智能视频的人数统计的研究与应用

发布时间：2018-03-03 16:06

本文选题：运动检测　切入点：目标跟踪　出处：《广西师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：近年来,随着电子信息产业制造技术的提高、硬件成本的降低,给计算机视觉技术带来了飞速发展的机会,加上社会各行业对智能视频的需求与日俱增,使得智能视觉成为一个理论和技术应用上热门的研究领域,基于视频的智能人数统计成为了该领域的热门研究方向之一。本文设计的基于视频的智能人数统计系统主要是针对于商场的人数统计,通过统计主要通道客流状态,从而进行店面的合理分布;统计各个区域的吸引率和繁忙度;有效评估所举行的营销和促销投资的回报;显示当前客流状态和变化趋势,安全保卫部门可以对流量较大的区域采取预防突发事件的措施,并可实时观察当前的实际人数及图像等,比传统的监控系统更加智能化。本文从硬件和软件部分进行设计,硬件方面与传统方案区别不大,只是比一般的系统增加了基于App的客户端,所以不做详细介绍,系统实现的重点与难点在于对运动人体目标检测技术、头部识别的机器学习方法和移动目标跟踪计数等,所以从以下几个方面展开研究:首先,分析人数统计系统使用的环境是商场的门口等人流密集的地方,通过方案比较,我们选取采用Rossi等提出的摄像机垂直架设拍照方法,采取检测人体头部区域。其优点是当行人靠近或肢体之间相互接触发生遮挡时,依然能够提取行人较为完整的头部信息,尽可能地减少由于遮挡引起的漏判。然后,使用帧间差分法从视频帧中提取运动目标,然后再以运动目标的外接矩形框作为后续人头检测的检测区域。研究表明行人头部检测采用基于HOG特征提取、线性支持向量机作为分类器的检测方法,是目前行人检测中综合性能较好的。接下来对视频场景中的行人进行跟踪计数。研究了经典的HS光流法和LK光流法。其中LK光流计算方法因为灵活性高、计算量相对较小更适合应用在目标跟踪中。对于空间运动位移较大的光流计算,将图像进行金字塔分解来提高光流矢量求解的精确度。在计数时,对场景设定感兴趣区域,并只对经过感兴趣区域的行人进行计数,并可以准确的判断进出方向。可以根据实际的需要情况,随意设定感兴趣区域,从而提高了系统的实用性。最后,对于APP的开发,本文主要介绍iOS操作系统的APP客户端的开发,采用的流程为:服务器端把检测到人数变化的图片保存为jpg格式文件,并存储在服务器,再向远程控制终端发送通知,控制终端解析推送通知,通过协议请求服务器中的图片以获得人数统计结果的实时图片。本文是以运动目标前景检测、基于机器学习的头部识别以及目标跟踪等技术在人数统计系统中实现了具体的应用案例。并且能够成功地在iOS手机客户端接收到人数变化的通知,得到人数变化时的图片。为了验证所用到的算法在本文提出的硬件配置要求不高的系统中的有效性、实时性及可靠性,采取了对大量不同场景下及不同人数的条件下的视频进行了测试,测试结果表明系统对于人数统计能够准确、有效的检测视频当中的行人头部,并在跟踪计数时具有较好的实用效果,达到预期设计目的。
[Abstract]:In recent years, along with the electronic information industry of manufacturing technology, reduce the cost of hardware, has brought great opportunities for the development of computer vision technology, with all sectors of society, demand for intelligent video makes intelligent vision become grow with each passing day, a theory and technology applied on the hot research field, the number of intelligent video based on statistics has become one of the most popular research direction in this field. The design of the intelligent video system based on the number of statistics is mainly based on the number of shopping malls statistics, through the statistics of the main channel flow state, so as to store the reasonable distribution; statistics in various regions of the attractive rate and busy degree; effective evaluation of a marketing and promotion investment return; display the current status and trend of the passenger flow, the security departments can take emergency prevention measures for regional heavy traffic, and real-time observation The actual number and image, more intelligent than the traditional monitoring system. This paper designed from hardware and software, hardware and the traditional scheme is very different, than the average increase of App system based on client, so the details do not system, emphases and difficulties of the realization of human motion target detection the head of the recognition technology, machine learning method and moving target tracking and counting, so from the following several aspects: first, analysis of the use of statistical system environment is the mall entrance and other populated areas, through the plan comparison, we selected by Rossi's camera is vertically erected photographing method detected by human head region the utility model has the advantages of contact with each other. When the occlusion occurred between pedestrians or near the limb, still be able to complete the extraction of pedestrian head information, as far as possible To reduce the occlusion caused by the leakage judgment. Then, using the frame difference method to extract moving objects from video frames, then the moving target rectangle as the detection area following head detection. According to the research on pedestrian head detection using HOG based feature extraction, linear support vector machine classifier as detection method at present, pedestrian detection is a good comprehensive performance. The next track count of the pedestrian in the video scene. The classic HS LK optical flow method and optical flow method. The LK optical flow calculation method for high flexibility, less calculation is more suitable for application in target tracking. The space motion of large displacement optical flow calculation. The image of the Pyramid decomposition to improve the accuracy of optical flow vector solution. In the count, area of interest to the scene, and only to the region of interest for pedestrians The number, and can accurately determine the direction of import. According to the actual situation, arbitrarily set the region of interest, so as to improve the practicality of the system. Finally, for the development of APP, this paper mainly introduces the development of iOS operating system APP client, the server process is: to detect the number of pictures saved as JPG files, and stored in the server, and then sent to the remote control terminal, the control terminal of push notifications, real-time image server in the picture by protocol request to obtain statistical results. This paper is based on the number of motion object detection, machine learning head recognition and target tracking technology to achieve the application the specific cases in the statistical system based on. And successfully received at the mobile phone iOS client to change the number of notification, the picture changes. In order to get the number of inspection The effectiveness of the system hardware configuration requirements proposed by the algorithm used in this paper is not high in the real time and reliability, take on a large number of different scenarios and different number of video conditions were tested, the test results show that the system can accurately for the number of statistics, the effective detection of pedestrian head video, and has a better practical effect in the tracking and counting, achieve the expected design objective.

【学位授予单位】：广西师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】