基于运动历史图像的异常行为识别算法研究

Research on abnormal behavior recognition algorithm based on motion history images

针对日益突出的安全监护问题,本文建立了一个能够较好地识别和预测人体摔倒等异常行为的算法模型。首先,通过人体动作实验分析人体基本的动作特征,运用背景减和混合高斯背景提取算法分别求解人体静态和动态前景,获得人体运动历史图像,运用运动历史图像进行运动分割,并通过Sobel算子求梯度判定人体运动方向,结合运动角度将得到的特征与训练出的动作库里的动作进行特征匹配,进而判断人体行为是否异常。通过这种方法构建的视频监控系统兼具实时性和智能性,可以为视频监控系统的开发提供理论依据。

To solve the increasingly prominent safety monitoring problem,an algorithm model which can effectively identify and predict abnormal behaviors such as human falls is proposed in this study.Firstly,through human motion experiments,the basic motion features of the human body are analyzed.Background subtraction and mixed Gaussian background extraction algorithms are used to separate the static and dynamic foreground of the human body respectively to obtain human motion history images.Motion segmentation is performed with motion history images,and the direction of human motion is determined through Sobel algorithm gradient.The obtained features are matched with the trained actions in the action library based on the motion angle.Then abnormal human behavior is recognized.The video surveillance system constructed using this method possesses real-time and intelligent capabilities,providing a theoretical basis for the development of video surveillance systems.