【作者】 徐光宇；

【导师】 崔荣一；

【作者基本信息】 延边大学 ， 计算机应用技术， 2011， 硕士

【摘要】 基于视觉的人体行为分析是当前一个非常活跃的研究领域,而对人的行为进行理解和分析则是该领域内富有挑战性的研究方向,它在智能监控、感知接口和基于内容的视频检索等领域具有广泛的应用前景。目前,人体异常行为的研究主要针对单人典型行为进行,而对单人潜在的异常行为及多人冲突行为的研究较少,而且行为分析还没有一个系统的理论框架。作为行为分析基础的运动目标检测算法因受光线变化等噪声的影响比较大,导致行为分析的效果不理想。本学位论文提出了利用人体上半身质心轨迹分析室内异常行为的方法,并提出了利用能量来分析两人间冲突行为以及利用颜色模型来检测吸烟行为的方法。首先,在室内环境下,通过自适应背景更新方法提取出运动目标并划分连通区域,根据连通区域的个数将该场景划为单人或多人。在单人场景下,提取运动目标上半身质心运动轨迹,根据其周期特征分析该人行为是否正常。其次,在多人无遮挡场景下,通过R变换识别个体动作,从而判别异常动作。在多人相互遮挡的场景下提取运动目标的运动能量,根据其能量图分析是否发生冲突行为。最后,在特定场景下,通过肤色模型判断场景中是否存在人体,在有人的前提下,结合火焰在时域中亮度瞬间变化和烟雾由下至上的运动特征,判断该场景中是否存在吸烟行为。仿真实验结果表明,基于人体上半身质心运动轨迹的行走状态分析方法能够准确分析室内行走的运动状态,识别率达到92%,并且对光线变化、背景噪声具有良好的鲁棒性。运动能量能较好地识别单人潜在的异常动作及多人冲突行为,识别率分别达到84%和80%。结合多特征的吸烟行为检测方法,能够较好地判断场景中是否存在吸烟行为。更多还原

【Abstract】 Vision-based human behavior analysis is currently one of the most active research fields, and to understand and analyze the human behavior is a challenging research topic. Human behavior analysis has many promising applications such as intelligent surveillance, perceptual interface and content-based video retrieval. Most research on human behavior analysis is based on single person and simple action; however, there is less research about single potential abnormal behavior and multi-person fighting behavior. As the basic of the human behavior analysis, the result of human behavior analysis always gets poor effect because of the noise in motion detection, such as varying luminance. In this dissertation, an approach to analyzing person states based on multi-viewing-angle was proposed, and then, a method to analyzing fighting behavior based on energy and to detect smoking based on color model were also proposed.Firstly, the moving person was detected using self-adaptive background updating algorithm and scene was estimated by the means of connected regions. After searching the centroid of upper part of the body, its periodic feature was extracted and the behavior was analyzed in single person scene.Secondly, human motion was discriminated by R transform in scene where there was no occlusion, and then fighting behavior was analyzed based on energy image in multi-person scene.Finally, in a given scene, person was detected by detecting skin area. And smoking was detected and analyzed by varying luminance, color model and directional information.The experimental results show that the human speed and walking states method is reasonably robust in varying luminance and shadow, and the accuracy of walking states is 92%. Moving energy can be used to recognize abnormal behavior, and the accuracy is 84% and 80% in single person and multi-person scene respectively. Smoking behaviors can be exactly analysis by multi features.更多还原