视频监控中人体行为识别的研究

Study of human behavior recognition based on video surveillance

提出一种人体行为识别模型和前景提取方法。针对人体运动过程中产生新的行为问题，该模型用分层Dirichlet过程聚类人体特征数据来判断人体运动过程中是否有未知的人体行为模式；用无限隐Markov模型对含有未知行为模式的特征向量进行行为模式的有监督的学习，由管理者将其添加到规则与知识库中。当知识库的行为模式达到一定规模时．系统便可以无监督地对人体行为进行分析，其分析采用Markov模型中高效的Viterbi解码算法来完成。对于前景的提取，提出了基于背景边缘模型与背景模型相结合的前景检测方法，此方法能够有效避免光照、阴影等外部因素的影响。仿真实验证明，本文提出的方法在实时视频监控中的人体行为识别方面有独特的优势。

A behavior recognition model and a model-based approach to extract foreground are presented. Because the process of human motion are likely to result in new events and behaviors, in this model, hierarchical Dirichlet process is used to cluster feature data monitored of the human body to determine whether there are unknown behaviors. Infinite hidden Markov model is used to learn unknown behavior patterns with supervise, and then update the knowledge base and rule base. When the rules and knowledge repository reach a certain scale, system can analyze behaviors without supervise. The Viterbi decoding algorithm in Markov model is adopted to analyze the current behavior of the human body. Foreground detection method is proposed based on the background edge model combined with background model, which can effectively avoid the light, shadows and other external factors. The simulation experiments show that this method in real-time video surveillance in the detection of human behavior has an unique advantage.