【作者】 张敏；

【导师】 宋井宽；

【作者基本信息】 电子科技大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 随着物联网技术的快速发展,由视频监控技术为主要载体的监控网络正在呈现爆发式的增长,以“天网系统”为代表的视频监控系统已经广泛应用于安全预警、商业经营、打击违法犯罪等领域。然而,由于视频监控系统体量庞大、数据量众多,信息维护十分困难。当前,对视频监控系统的数据处理主要以人工维护为主,资源成本耗费量大,同时容易产生工作效率低下、信息利用率低和误差率高等问题。因此,亟需一套能够对视频进行自动化处理的智能视频分析处理系统,该系统通过计算机的高速分析处理能力,把需要人工去完成的工作转变为让计算机去完成的工作,在对海量视频监控数据分析处理的过程中,有效提取系统使用者需要关注的重点信息,最大程度地避免人工劳动带来的误差,从而完美实现“机器替人”的工作。随着随着深度学习技术的不断发展,机器能够对图像有更深层次的理解能力,深度神经网络在图像识别领域已经超过了人类,并在目标检测等领域远胜于其他传统方法。因此,拟利用深度神经网络^[1]的强大图像处理能力对海量视频监控数据进行分析处理,从而实现对视频监控数据的有效处理和有用信息的更深层次的挖掘。本文研究内容主要围绕室内复杂环境展开,针对当前视频监控处理技术存在的矛盾问题,拟从计算机视觉的角度,深入剖析研究室内复杂空间环境下目标检测识别^[2]和多目标跟踪^[3]等问题。通过研发出室内复杂环境下人员识别跟踪系统,实现对人员实时定位和行为轨迹跟踪,进而实时了解人员动态,精准掌握人员活动轨迹。本文主要研究室内复杂环境下人员检测识别和轨迹跟踪,通过对研究目的的梳理和分析,发现达到研究目的过程中有几个需要解决的难点问题:目标遮挡问题、目标丢失问题、复杂背景问题、多视角问题问题、复杂场景下的目标检测和识别问题,以及多视角切遮挡频繁情况下的目标跟踪问题。其中遮挡问题、目标丢失问题、复杂背景问题、多视角问题这几个问题都可以归为多视角且遮挡频繁情况下的目标跟踪问题。整个方法的简要流程如下:对监控视频中的帧处理提取图像序列,使用Center Net基于目标关键点的检测算法与轨迹跟踪算法提取图像序列中的目标,采用“特征提取记录+数据库查询+目标身份识别”的方法实现对系统中待检测目标（人）的身份识别,随后使用匈牙利算法和卡尔曼滤波完成多目标跟踪步骤,进一步能够得到进入室内人员的轨迹路线。这样通过人工智能的技术,对室内人员进行径迹的实时捕捉,完成智能化地对进入室内人员的相关信息的捕捉、跟踪、统计和记录,最终实现室内复杂环境下人员识别跟踪系统的研发。通过系统测试及运用,室内复杂环境下人员识别跟踪系统可以达到实时准确的目标检测、目标识别和多目标跟踪效果,为系统正确决策提供强有力的数据支持。更多还原

【Abstract】 With the rapid development of Internet of things technology,the monitoring network with video surveillance technology as the main carrier is showing explosive growth.The video surveillance system represented by "Skynet system" has been widely used in the fields of security early warning,commercial operation,combating crimes and so on.However,due to the huge volume and large amount of data of video surveillance system,information maintenance is very difficult.At present,the cost of video surveillance and data processing is high,and it is easy to produce errors due to the high utilization rate of video resources.Therefore,there is an urgent need for an intelligent video analysis and processing system that can automatically process video.Through the high-speed analysis and processing ability of the computer,the system transforms the work that needs to be completed manually into the work that needs to be completed by the computer,and effectively extracts the key information that system users need to pay attention to in the process of analyzing and processing massive video surveillance data,Avoid the error caused by manual labor to the greatest extent,so as to perfectly realize the work of "machine replacing people".With the continuous development of deep learning technology,machines can have a deeper understanding of images.Deep neural network has surpassed human beings in the field of image recognition,and is far better than other traditional methods in the field of target detection.Therefore,it is proposed to analyze and process massive video surveillance data by using the powerful image processing ability of deep neural network,so as to realize the effective processing of video surveillance data and deeper mining of useful information.The research content of this thesis mainly focuses on the indoor complex environment.In view of the contradictions existing in the current video surveillance and processing technology,this thesis intends to deeply analyze and study the problems of target detection and recognition and multi-target tracking in the indoor complex space environment from the perspective of computer vision.Through the research and development of personnel identification and tracking system in indoor complex environment,real-time personnel positioning and behavior track tracking are realized,so as to understand personnel dynamics in real time and accurately grasp personnel activity track.This thesis mainly studies the detection,recognition and trajectory tracking of people in indoor complex environment.Through the combing and analysis of the research purpose,it is found that there are several difficult problems to be solved in the process of achieving the research purpose: target occlusion,target loss,complex background,multiperspective,target detection and recognition in complex scenes,And the target tracking problem in the case of frequent multi-view cut occlusion.Among them,occlusion problem,target loss problem,complex background problem and multi-view problem can be classified as target tracking problem in the case of multi-view and frequent occlusion.The brief flow of the whole method is as follows: extract the image sequence from the frame processing in the surveillance video,extract the target in the image sequence using centernet’s detection algorithm based on target key points and trajectory tracking algorithm,and realize the identification of the target(person)to be detected in the system by using the method of "feature extraction record + database query + target identification",Then the Hungarian algorithm and Kalman filter are used to complete the multi-target tracking step,and the trajectory of people entering the room can be further obtained.In this way,through the technology of artificial intelligence,the track of indoor personnel can be captured in real time,and the relevant information of indoor personnel can be captured,tracked,counted and recorded intelligently,so as to finally realize the research and development of personnel identification and tracking system in indoor complex environment.Through system testing and application,the personnel identification and tracking system in indoor complex environment can achieve real-time and accurate target detection,target identification and multi-target tracking,and provide strong data support for the correct decision-making of the system.更多还原