公共场合打架和抢劫行为识别技术研究

发布时间：2018-07-18 11:19

【摘要】：目前在海量监控视频中发生的打架斗殴以及抢劫等恶劣行为,人工用眼睛无法对视频进行长时间持续性的观测,如何实时的对监控视频中的异常行为进行检测和识别是智能安防领域不断研究的课题。本文主要研究复杂背景环境和多人流环境中运动人体的异常行为识别,研究内容分为目标的有效跟踪和异常行为识别两点:一、通过混合高斯背景建模的背景差分法实现视频下运动人体快速检测,实时记录每个目标的运动轨迹,当目标正常行走时采用经验距离结合分块直方图匹配的方法进行跟踪;当目标出现遮挡、交汇和丢失的情况时,采用卡尔曼滤波结合区域光流角点跟踪算法,对目标运动的位置进行预测的同时利用光流角点的微观跟踪进行位置校正,实现每个目标的有效跟踪。二、通过区域加权光流直方图初步确定可能打架的目标,再使用SVM分类器结合HOG特征训练的上半身攻击行为分类器检测是否具有打架动作,最后通过加权梯度直方图对检测到的出拳胳膊位置计算直方图变化率,变化频率高于阈值且持续20帧以上则识别为打架行为;通过搭建异常速度模型,比对区域光流异常的位置信息,结合卡尔曼滤波算法进行抢劫行为的识别。最后通过多次实验证明,上述方法可以对打架和抢劫这两项异常行为进行有效的识别,准确率较高。
[Abstract]:At present, in the mass surveillance video, the bad behaviors such as fights and robberies, can not be observed continuously for a long time with artificial eyes. How to detect and identify abnormal behavior in surveillance video in real time is a research topic in the field of intelligent security. This paper mainly studies the recognition of abnormal behavior of moving human body in complex background environment and multi-human flow environment. The research content is divided into two aspects: effective tracking of target and recognition of abnormal behavior. The background differential method of hybrid Gao Si background modeling is used to realize the fast detection of moving human body under video, and the moving track of each target is recorded in real time. When the target is walking normally, the method of empirical distance and block histogram matching is used to track the target. When the target appears occlusion, intersection and loss, the Kalman filter combined with the regional optical flow corner tracking algorithm is used to predict the moving position of the target, and the micro-tracking of the optical flow corner is used to correct the position of the target. Achieve effective tracking of each target. Secondly, by using the area weighted optical flow histogram to determine the possible target of fighting, and then using SVM classifier combined with HOG feature training of the upper body attack behavior classifier to detect whether there is a fight action, Finally, the change rate of histogram is calculated by weighted gradient histogram. The frequency of change is higher than threshold and lasts more than 20 frames, then it is recognized as fighting behavior. Comparing the location information of regional optical flow anomaly, the robbery behavior is identified by Kalman filter algorithm. Finally, it is proved by many experiments that the above method can effectively identify the two abnormal behaviors of fighting and robbery, and the accuracy is high.
【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】