【作者】 杨硕；

【导师】 刘富；

【作者基本信息】 吉林大学 ， 工程硕士（专业学位）， 2017， 硕士

【摘要】 随着监控设备硬件的成熟和计算机视觉相关技术的进步,监控设备在安全防控、交通管理、刑事侦查等多个领域得到了广泛应用。一方面是全世界都面临的严峻反恐形势,另一方面是民众不断增强的安全防控意识,在这样的双重背景下,视频监控正在发挥着越来越重要的作用。但是视频文件包含的信息量非常巨大,如何进行准确、高效地处理和分析是一个难题。在公安刑侦领域,目前对于监控视频中嫌疑人目标的追踪大都依靠人力来完成,其效率非常低。现有的监控视频检索方法在效率和准确率上表现较差。针对上述问题,我们迫切地需要新的技术手段来解决,希望能够实现更准确、更快速地在监控视频中查找相关目标。为了解决这些问题,本文深入研究了基于内容的监控视频检索的相关技术,具体来说,本文的主要研究工作如下:(1)为了解决传统关键帧及关键帧ROI提取算法的准确度普遍较低的问题,本文提出了一种基于人体检测的关键帧及ROI提取算法。该算法通过引入人体检测技术,实现了更准确地选取关键帧以及更精准地标定关键帧ROI区域。实验表明,该方法能够更准确地提取关键帧和关键帧ROI。(2)现有的人体检测算法在复杂场景下的检测效果并不理想,会造成嫌疑人目标的漏检。为了改善在复杂场景下的人体检测效果,本文提出了多部位最优特征组合的人体检测算法,提高了复杂场景下人体检测的准确率。(3)深入研究了关键帧检索匹配算法,针对传统SIFT算法存在的缺陷,本文采用有监督的LPP算法和基于向量夹角的近似最近邻匹配算法对SIFT特征进行降维和加速匹配,有效提高了SIFT特征点的匹配速度。为了提高检索的准确率,本文提出了一种融合颜色特征的改进SIFT算法。通过实验,验证了该算法的检索准确率更高。(4)本文设计并实现了一个监控视频检索云平台,主要包括云平台的构建、图像处理模块和数据库模块的开发部署等工作。该平台可以提供稳定、高效、准确的检索服务,并生成检索报告为刑侦人员提供参考。通过性能测试实验,验证了该平台的功能稳定而实用。更多还原

【Abstract】 With the development of video surveillance equipment and the computer vision,video surveillance has been widely used in the security prevention,traffic control,criminal investigation and other fields.On the one hand is the world is facing a serious anti-terrorism situation,on the other hand is the people’s growing awareness of security,in such a dual context,video surveillance is playing an increasingly important role.But the video data contains a very large amount of information,it’s a big problem to processing the massive video data accurately and efficiently.In the field of public security criminal investigation,the current way of the suspects tracking in the video mostly rely on manpower to complete,its efficiency is very low.And the existing surveillance video retrieval method has a bad performance in the efficiency and accuracy.In view of the above problems,we urgently need new technical means to solve,hoping to achieve more accurate and faster in the surveillance video to find relevant targets.In order to solve these problems,this paper makes a deep research on the key technology of content-based surveillance video retrieval.Specifically,the main research work of this paper is as follows:(1)Aiming at the problem that the accuracy and efficiency of the traditional key frame and key frame ROI extraction algorithm are generally not high,this paper presents a key frame and ROI extraction algorithm based on human detection.By introducing human detection technology,this algorithm realizes more accurate selection of key frames and more accurate calibration of ROI regions.Experiments show that the method has a better effect.(2)The existing algorithm’s detection effect of human body detection in complex scene is not ideal.In order to improve the detection effect of human in complex scenes,this paper presents a multi-part optimal feature combination of human detection method to improve the accuracy of human detection in complex scenes.(3)In this paper,we focus on the key frame retrieval matching algorithm.In view of the shortcomings of the traditional SIFT algorithm,this paper adopts the supervised LPP algorithm and the approximate nearest neighbor matching algorithm based on the vector angle to reduce the dimension and accelerate the SIFT feature’s matching,and effectively improve the SIFT feature point matching speed.In order to improve the accuracy of retrieval,this paper presents an improved SIFT algorithm that combines color features.The experimental results show that the algorithm has higher retrieval accuracy.(4)This paper designs and implements a surveillance video retrieval cloud platform,including the construction of cloud platform,image processing module and database module development and deployment work.The platform can provide stability,colleges and universities,accurate retrieval services,through the performance test,verify the effectiveness and practicality of the platform.更多还原