基于CNN特征与HMM时序建模的人体行为识别研究
首发时间:2018-03-08
摘要:随着视频的总量和类别数量迅速增长,智能化的视频分析方法已经成为许多实际应用中的核心技术,例如异常监控、医疗诊断和视频检索等等。人是社会的主体,所以视频中的大多数都是关于人体行为的内容,对人体行为的识别也就成为视频分析的核心技术。现阶段对图像的识别技术已经成熟,而且视频可以看作是一个图像序列,所以对视频的分析可以转换成图像的处理。深度学习是图像处理方面现阶段最为火热的方法,其中的CNN模型已经取得了非常好的成果,相较于一些传统的人工特征提取方法有着很大的优势。视频是具有时序信息的信号,所以在分类识别时更好的利用时序信息能够大大的提升识别准确率。本文提出一种人体行为识别方法:利用CNN对视频进行图像特征提取,然后应用HMM对图像特征序列进行时序建模,创新性的使用HMM的参数作为视频整体的特征表示。最后,本文的方法在公共人体行为数据库上进行了验证,取得很高的识别准确率。
For information in English, please click here
Research on human behavior recognition based on CNN features and HMM time series modeling
Abstract:With the rapid growth of the total number and category of video, intelligent video analysis has become the core technology in many practical applications, such as anomaly monitoring, medical diagnosis and video retrieval. People are the main body of the society, so most of the videos are about the contents of human behavior, and the recognition of human behavior has become the core technology of video analysis. At present, the technology of image recognition is mature, and video can be regarded as an image sequence, so the analysis of video can be converted into image processing. Deep learning is the most popular method in the field of image processing. The CNN model has achieved very good results, which has great advantages compared with some traditional manual feature extraction methods. Video is a signal with time sequence information, so better use of time sequence information in classification recognition can greatly improve the accuracy of recognition. In this paper, a human behavior recognition method is proposed, which uses CNN to extract the image features, and then applies HMM to sequence the image feature sequence, and innovatively uses the parameters of HMM as the overall feature representation of the video. Finally, the method of this paper is verified on the public human behavior database, and the accuracy of recognition is very high.
Keywords: action recognition CNN feature extraction HMM feature representation
基金:
引用
No.****
动态公开评议
共计0人参与
勘误表
基于CNN特征与HMM时序建模的人体行为识别研究
评论
全部评论