Human Behavior Recognition Based On Local Space-time Video Features

In the era of the rapid development of computer and network technology,multimedia information has penetrated into all aspects of the daily production and life.Massive growth of data,frequent interaction and migration to cloud service are the three characteristics of the advance of multimedia technology.Therefore,it is a very pressing task to quickly search for videos that users are interested in or to quickly categorize videos for afterward date processing.Human behavior recognition is an important part of the computer video field,in recent years,it has made a great contribution to the research of practical problems,such as intelligent security,video retrieval and human-computer interaction.Human behavior recognition method based on local space-time features is the current mainstream method,which can aggregate the local low-level features and establish the associate to the high-level semantics,and has become a standard paradigm for human behavior recognition.Firstly,it needs to extract features from the video;and then uses feature coding method to describe the video with histogram vector;finally,selects the appropriate classifier to identify the behavior.So,the video feature extraction is very critical.To overcome the gap between the low-level features and the high-level semantics,combining with the development of human behavior recognition technology,this paper mainly does the following work:1.Systematically summarize existing human behavior recognition methods.Access to a large number of related literatures,existing human behavior recognition methods based on local space-time features are outlined.The commonly used space-time feature descriptors,feature coding methods and Bag of Words(BoW)model have been illustrated in detail.2.We propose the human behavior recognition algorithm based on locality and collaborative representation.In order to make full use of video data and ensure the highly descriptive histogram vector of behavior,a novel and effective algorithm for human behavior recognition is proposed.The proposed algorithm combining collaborative representation with space-time pyramid,which preserves the similarity between the test sample and its neighboring training samples by employing the locality.Therefore,the algorithm not only avoids the dilemma of expensive computations in the sparse representation,but also improves the discriminative power of histogram vector by encoding the structured distribution of features in the video.3.We propose an algorithm of human behavior recognition based on hash codes of space-time video features.The process of learning auto-encoder hashing functions and mapping local feature points into binary hash codes is focused.The method changes the coding mode of the traditional BoW model,which uses binary hash codes to describe the local features and preserves the local similarity information of a feature;then,on the binary hash codes,it performs K-Means cluster.Therefore,many powerful discriminative visual dictionaries are generated.Finally,combining with the space-time pyramid model,the video is expressed as a histogram vector of space-time pyramid.Comparing with the traditional method of human behavior recognition based on sparse coding and its improved algorithm,Experimental results show that the proposed algorithm had shorter time of learning vocabulary,faster encoder speed and higher accuracy of human behavior recognition.4.Finally,we make a summary and put forward some prospects for further research.