Research On The Technologies For Human Action Recognition

Action is the set of organism reactions to adapt to a variety of changes from physical and external environment. Action recognition is a hot spot in computer vision, behavioral psychology and so on. Human action recognition is one important direction of action recognition. and it aims to reveal the people’s inner intentions. Video-based human action recognition is an important branch of human action recognition. It shows the great potential economic and social value through the widely applications in remote control system, intelligent traffic, robotic, video surveillance, action analysis, human-computer interaction, etc. Since the research involves various kinds of knowledge, such as artificial intelligence, pattern recognition, psychology, image processing, and human description, there are still many technologies need to break. This thesis sets to study the key technologies of video-based action recognition, including foreground extraction, feature extraction and selection, and the design of classification algorithm. The main contribution of this thesis can be concluded as follows.(1) Considering that the existing test database is of low resolution while most of the existing digital equipment is of high resolution, which results in unsatisfied performance of the original algorithm. We construct a high resolution database of the human action videos captured by one Sony cameras indoor. Its frame rate is30frames/second, and its resolution is1440*1080. The database contains four types of human actions:squat, standing up, waving, fall.(2) Considering that the common morphological methods usually cause distortion and can’t get a good effect for filtering the large filed of noise in the segment of foreground extraction, this thesis proposed an de-noising algorithm based on hierarchical filters. The algorithm define independent connected region, build searching rules, and filter noise using the screened ideas from coarse to fine based on statistical difference of foreground and background. Experimental results show that the proposed method has a better result and reduce distortion.(3) For the technologies of feature extraction and selection, this thesis proposes a new global feature model called action accumulation feature vector (AAFV). The model firstly define an accumulation action image based on Motion Energy Image, then get a global vector which retains the spatial and temporal information after divided into grids and analysis in frequency domain.(4) For the technologies of classification algorithm design, this thesis proposes two new algorithms of human action recognition. In the first algorithm, the human action are clustered based on Motion Energy Image(MEI), then a hierarchical recognition method is designed by AAFV and local features using Support Vector Machine (SVM). The second algorithm proposed the mean AAFV of different kinds of human actions as the pre-judgment to select the high probability actions against the complexity and diversity of human actions. Then the high probability actions will be classified using the AAFV and local features. The experiment results show that the proposed method can recognize the fine action, reduce the complexity of the classification algorithm and improve the recognition efficiency.