A Research On Methods Of Human Activity Recognition

Human activity recognition in videos is a research aimed to identify human behaviors correctly, by means of detecting and describing the features of the moving human body in the collected video sequences, constructing an effective classification and recognition model and automatically analyzing and understanding the video contents.In this thesis, the main contents are as follows:(1) choose the efficient robust space-time interest points as the features for human activity recognition research.On the basis of analyzing and comparing several methods of human behavior detection, choose Dollar’s method whose performance is efficient and robust to detect the space-time interest points of human activity in this paper.(2) Present two improvements on feature description for space-time interest points.① Smoothing the encoding threshold by combining with the neighborhood information to reduce the sensitivity of LBP algorithm to noise, while keeping effective feature information. The experimental results show that the proposed improvement highlights the local features and improves the human behavior recognition accuracy.② LBP-TOP algorithm ignores the relationship between adjacent points which easily causes a loss of neighborhood structure information, and neglects the differences of changes among the local features of space-time interest points between spatial and time domains. To overcome the shortcomings mentioned above, propose ST-LBP-TOP algorithm which uses an unequal strategy to describe the local features of space-time interest points. Use Direction coded Local Binary Pattern in both spatial and temporal domains to obtain the texture of neighborhood structure. In addition, introduce Linear Local Binary Pattern to describe the texture changes along the temporal domain.(3) Propose an improved BOW model with double dictionaries and a dictionary optimization strategy.① To solve the problem of BOW model neglecting the local features relationship between spatial and temporal domains, present a model with double dictionaries based on local features’ apparent information and space-time locations information respectively, so that enrich the semantic information of dictionaries. The experimental results show that the improved BOW model with double dictionaries has better classification accuracy.② To reduce the quantization errors using K-means clustering method to generate dictionaries, propose to assign a Local feature to multiple words according to different weights of membership, in which way introduce K-SVD algorithm to generate and optimize the dictionaries. The experimental results show that the improved BOW model and the optimizing method for dictionaries generation achieve better accuracy of human activity recognition.