基于视频图像的人体异常行为识别技术研究

发布时间：2018-03-05 14:22

本文选题：人体异常行为识别　切入点：Kalman滤波　出处：《哈尔滨工程大学》2012年硕士论文　论文类型：学位论文

【摘要】：人体异常行为识别是对监控范围内的人体行为进行分析，将其与预先设定的异常行为进行匹配，如果行为匹配，则判定为异常行为并自动报警。传统的监控系统的功能单一，主要依靠监控人员人为操作来监控短而随机的突发和异常事件，要求监控人员持续地监视屏幕，不仅需要大量的监控人员，而且容易因为监控人员的松懈、监控点较多无法做到实时完整监控而导致漏警，造成不可估量的损失。研究内容包括：首先，以背景减除法为基础，对混合高斯模型进行改进并将其作为背景模型。对混合高斯模型参数均值和方差采用不同的学习率，用以提高混合高斯模型背景提取的准确程度。将马尔可夫随机场理论引入前景分割中，以此提高前景分割精度，并提出一种方法使得人体运动静止时也能够有效检测。针对HSV颜色空间在阴影检测方面的优越性，结合上面的检测方法，提出一种能够准确快速消除阴影的方法；其次，为了能够实时自动跟踪并且不受相互遮挡的影响，将Mean shift算法和Kalman滤波算法进行融合，提出新的算法，预先获得人体运动信息，，并实现快速搜索；最后，提出一种特征提取方法，使用HMM进行训练和分类，用以识别人体异常行为。通过对所提出的方法进行实验验证，确定该方法可以有效的识别晕倒、砸车等人体异常行为。
[Abstract]:Human abnormal behavior identification is to analyze the human behavior in the monitoring area, and match it with the predefined abnormal behavior. If the behavior matches, it will be judged as abnormal behavior and alarm automatically. Mainly relying on the manual operation of the monitoring personnel to monitor short and random emergencies and abnormal events, requiring the monitoring personnel to continuously monitor the screen, not only requires a large number of surveillance personnel, but also easily because of the lax monitoring personnel. More monitoring points can not achieve real-time complete monitoring, resulting in false alarm, resulting in incalculable losses. The main contents are as follows: firstly, based on the background subtraction method, the mixed Gao Si model is improved and used as the background model, and the parameter mean and variance of the mixed Gao Si model are studied at different learning rates. In order to improve the accuracy of background extraction of mixed Gao Si model, Markov random field theory is introduced into foreground segmentation to improve the precision of foreground segmentation. In view of the superiority of HSV color space in shadow detection, combined with the above detection method, a method to eliminate shadow accurately and quickly is proposed. In order to track human body automatically in real time and without the influence of mutual occlusion, the Mean shift algorithm and Kalman filter algorithm are fused, and a new algorithm is proposed to obtain human body motion information in advance and to realize fast search. A feature extraction method is proposed, which is trained and classified by HMM to identify abnormal human behavior. Through the experimental verification of the proposed method, it is determined that this method can effectively identify the abnormal behaviors of human body such as fainting, smashing and so on.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP391.41

【参考文献】