基于加权HU不变矩的监控视频人体行为识别方法的研究与实现

【作者】 张琳；

【导师】 周明全；

【作者基本信息】 西北大学 ， 计算机技术， 2011， 硕士

【摘要】 随着视频监控系统在社区安防、机场安检、景区安全等方面的广泛应用,如何实现快速、精确的异常行为检测成为视频监控中一个亟待解决的问题,同时也成为近年来计算机视觉领域的一个研究热点。本文研究并改进了运动人体行为识别处理中的关键算法,主要工作进展包括：(1)在图形图像处理方面,针对监控原始图像模糊粗糙问题,本文采用对视频监控图像实施平滑和滤波处理,获得了清晰准确的改进图像。(2)在运动目标检测方面,本文提出背景减除法与时间差分法相结合的加权平均法。为保证检测的高效性、准确性,针对背景变化简单的情况,本文提出基于背景更新与自适应阈值选取相结合的方法,提高算法性能；针对背景复杂情况,本文提出基于高斯混合背景模型的运动检测方法,它根据实验结果分析动态调整混合高斯模型参数,有效解决光照以及树木扰动等外界变化带来的影响。(3)在运动目标行为识别方面,针对要提高人体异常行为识别效率,提出了一种以改进的加权Hu不变矩作为目标特征,利用滑窗机制对连续视频帧的匹配结果进行综合分析的方法,从而实现对人体异常行为的识别；同时研究了基于目标轮廓特征和基于隐马尔科夫模型的识别算法,以目标的最小外接矩形的长宽比、目标的覆盖度以及目标的轮廓特征组成的特征向量作为人体行为的特征描述,判别人体行为是否异常。本文设计实现了一个针对人体运动行为识别的智能视频监控系统,验证了改进的算法可有效检测出实时视频中运动的目标人体,并识别出其行为的变化。更多还原

【Abstract】 With the extensive use of video surveillance systems in Community security, safety and other aspects, How to achieve rapid,accurate detection of abnormal behavior become video monitoring a problem to be solved, but also become a hot research topic in the field of computer vision. This paper study and improved the movement Human behavior in the processing of identification key algorithms, mainly including work progress:(1)As to the image blurring rough for monitoring, before graphic image processing, we choose the methods of smoothing, filtering and other pre-processing, and then the image has not been destroyed.(2)In moving target detection, change the background image for the inconsistency, while ensuring the efficient simplicity, this paper presents a simple change in the background of the case that the problem of the shadow, and select updating and threshold selection combined to improve performance of the algorithm based on adaptive background; in the context of complex situations, we choose the Gaussian mixture model-based extraction of moving targets, through the analysis of test results, we improve the Gaussian mixture model parameters, and solve the light and the trees outside disturbances such as the impact of changes. Finally, a fast YUV color space shadow suppression algorithm.(3) In identifying the target behavior in the campaign, in order to make the efficiency for identifying abnormal behavior more convenient, we improve the weighted Hu invariant moments as the target, the use of sliding window mechanism to match the results of successive video frames a comprehensive analysis to human behavior identification; while the target contour based features and hidden Markov model-based algorithm for the minimum bounding rectangle of the target aspect ratio, target coverage and target contour feature vector consisting of features described as characteristic of human behavior, the human behavior on the basis of discrimination is abnormal.This design implements a video surveillance system of motion human identification. we show that:the proposed method can effectively detect moving targets in real-time video and identify the movement of human behavior.更多还原