基于视频监控的智慧幼儿园安全检测关键技术研究

发布时间：2018-02-03 23:08

本文关键词： 智慧幼儿园视频监控异常检测目标检测目标跟踪运动模型　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：基于视频监控的智慧幼儿园安全检测,是指以视频监控为载体对智慧幼儿园内幼儿的异常行为进行实时自动的检测,从而最大程度确保幼儿的安全。一方面,由于当今社会正处于信息化时代,物联网的存在使得我们可以获取到海量的数据;另一方面,随着人们生活水平的提高,家长越来越关注幼儿的成长和身心健康。相比于传统的幼儿园,智慧幼儿园的安全检测也需要更加完善,因此,本文致力于智慧幼儿园内的幼儿安全检测关键技术研究,以视频监控为载体,实现对智慧幼儿园的安全检测。按照视频监控场景的复杂度和异常检测的要求,园内幼儿安全检测可分为简单场景下的检测和复杂场景下的检测。首先,本文提出了越界检测、基于封闭区域内的规则检测和摔倒检测这三种基于规则的幼儿个体异常行为检测算法,主要用于园内简单场景下幼儿的异常行为检测。这三种检测算法实现简单,能够取得很好的检测效果,检出率分别为100%、100%和92%,但需要事先人为规定异常规则,可扩展性不强。其次,本文提出了一种基于改进的Camshift/Kalman个体运动多目标跟踪算法,并根据此跟踪算法进一步提出了一种基于运动模型的幼儿个体运动异常检测算法。其中,提出的个体运动多目标跟踪算法用于实现多目标跟踪、目标跟踪自动初始化、目标跟踪自动失效处理等功能。提出的幼儿个体运动异常检测算法在利用个体运动多目标跟踪算法获取到运动目标信息后,以视频监控图像中最小子块为单位,通过建立的正常运动模型对未知运动目标进行异常检测,实现对幼儿在视频监控中的位置、运动方向、运动速率和停留时间等信息的实时检测。该算法具有较高的准确性和较快的检测速度,特别是针对运动目标较少的场景中的个体运动异常检测,平均召回率为92.6%,平均检测速度为33.8 ms/帧,优于同类个体异常检测算法的实验结果数据。第三,本文提出了一种结合背景差分和光流法的幼儿群体运动状态突变事件检测算法,用于园内幼儿集体活动时群体运动状态突变的异常检测。该算法结合了背景差分算法数据处理速度快的优势和光流算法实验结果准确性高的优势,对于UMN视频数据库中的视频片段AUC值可以达到0.96以上,平均检测速度为25.8 ms/帧,相比于其它同类群体异常检测算法在综合性能上有明显的提高。最后,利用基于像素统计的人群密度估算方法,将以上提出的各种安全检测算法整合在一起,组合成一套比较完整的适用于智慧幼儿园内的视频监控安全检测算法,并利用OpenCV开源计算机视觉库,在 Windows 操作系统 Microsoft Visual Studio 2010 开发平台下用 Visual C++语言对智慧幼儿园视频监控安全检测关键技术进行了实现。
[Abstract]:Intelligent Kindergarten Security Detection based on Video Surveillance refers to the use of video surveillance as a carrier to carry out real-time and automatic detection of abnormal behavior of children in intelligent kindergartens so as to ensure the safety of children to the greatest extent. On the one hand. As the society is now in the information age, the existence of the Internet of things enables us to obtain a large amount of data; On the other hand, with the improvement of people's living standards, parents are paying more and more attention to the growth and physical and mental health of their children. This paper is devoted to the research of key technology of infant safety detection in intelligent kindergarten, and realizes the security detection of intelligent kindergarten with video surveillance as the carrier, according to the complexity of video surveillance scene and the requirements of abnormal detection. The safety detection of children in the garden can be divided into simple scene detection and complex scene detection. Firstly, this paper proposes the cross boundary detection. Rule detection and fall detection are three rule-based algorithms for individual abnormal behavior detection in closed areas. Mainly used for the detection of abnormal behavior of children in the simple scene of the garden. These three detection algorithms are simple and can achieve a good detection effect. The detection rate is 100% and 92% respectively. However, the outlier rules need to be stipulated in advance, so the scalability is not strong. Secondly, this paper proposes an improved multi-target tracking algorithm for individual motion based on Camshift/Kalman. According to this tracking algorithm, a motion model-based algorithm for individual motion anomaly detection is proposed, in which the proposed multi-target tracking algorithm is used to achieve multi-target tracking. Target tracking automatic initialization, target tracking automatic failure processing and other functions. The proposed individual motion anomaly detection algorithm after the use of individual motion multi-object tracking algorithm to obtain the information of moving targets. Take the smallest block in the video surveillance image as the unit, through the establishment of the normal motion model to detect the unknown moving object, realize the position and the movement direction of the child in the video surveillance. The algorithm has higher accuracy and faster detection speed, especially for individual motion anomaly detection in the scene with fewer moving targets. The average recall rate is 92.6 and the average detection speed is 33.8 Ms / frame, which is better than the experimental data of the same individual anomaly detection algorithm. Third. In this paper, an algorithm for detecting sudden events in children's movement state is proposed, which combines background difference and optical flow method. This algorithm combines the advantages of fast data processing of background difference algorithm and high accuracy of experimental results of optical flow algorithm. For UMN video database, the AUC value of video fragments can reach above 0.96, and the average detection speed is 25.8 Ms / frame. Compared with other similar groups of anomaly detection algorithm in the comprehensive performance is significantly improved. Finally, using the population density estimation method based on pixel statistics, the above proposed security detection algorithms are integrated together. A relatively complete set of video surveillance security detection algorithm suitable for intelligent kindergarten, and the use of OpenCV open source computer vision library. Using Visual C in Windows operating system Microsoft Visual Studio 2010. The key technology of security detection of intelligent kindergarten video surveillance is implemented in this paper.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN948.6

【参考文献】