Action Recognition Based On Spatial-Temporal Pyramid Sparse Coding

The research of human action recognition based on computer vision is anintersecting subject of Pattern Recognition, Computer Vision, Image Processing andso on. It has great application value and theoretical significance in human-interactive,object-based video retrieval, video motion analysis and intelligent video monitoring.It is the variety of human action, location of video camera and complex environmentthat makes the human action a challenging subject.This paper refers to a principal technique being used in contemporary work ofaction recognition, for example, support vector machine and bag of words whichraises an improved method of human action recognition, spatial pyramid with sparsecoding which applied in human action recognition of video. This paper expanded themethod of spatial pyramid matching into video analysis, accompanying with methodof sparse coding. With these two methods, we can describe a video from the time andspatial dimension, as well as recognize the action information according to spatialstructure and time sequence. The processes of the approach: first, abstract featurepoints from dense optical flow field, and generate the feature descriptor usingdisplacement information; second, generate the visual dictionary using sparse coding,here the dictionary is used to quantify the feature description, and the spatial-temporalpyramid model is used for videos; finally, classify the videos, and recognize thehuman actions.In the experiment part of this paper, we applied the classic Bag of Wordsframework. And in the training part, the support vector machine and Chi-Squarekernels used to train the classifier. In order to prove the validity of the method, we didplenty of experiments and the method performs well in classic action recognitiondatabase such as KTH, WEIZMANN and Hollywood.