基于HMM和LDA级联的视频异常检测

Video Abnormal Event Detection Based on HMM Cascaded with LDA

该文针对视频监控中的异常事件检测问题,提出了一种基于HMM和LDA的级联方法。该方法将提取的底层特征建立文本信息进行LDA训练后,得到的语义特征视为HMM观测量用于建立级联模型,从而对整体视频事件进行学习和异常检测。检测红灯禁行路段异常的实验结果表明,相对于基于HMM的方法,该方法在检测异常事件方面得到了更好的效果。

For video abnormal event detection, an approach based on HMM cascaded with LDA is proposed. Markov chain-based HMM describes dynamic sequence, and LDA can mine topic feature of events efficiently. The approach firstly extracts robust SIFT feature points from image sequences, and in order to train these feature points, then transforms them into text using LDA. Finally, HMM learns topic feature and detects abnormal events sequentially. Experiment results in red light road show that the approach has a better effect.