A Novel Action Recognition Scheme Based on Spatial-Temporal Pyramid Model

Recognizing actions is one of the most important challenges in computer vision. In this paper, we propose a novel action recognition scheme based on spatial-temporal pyramid model. Firstly, we extract the basic visual feature descriptors for each video. Secondly, we construct visual dictionary on the whole visual features set. Thirdly, we construct a novel spatial-temporal pyramid model by dividing the visual features set of each video into multi-scale blocks in 2-dimensional space domain and 1-dimensional time domain separately. Then we calculate the distribution histogram representation for each block of different scales by using the bag-of-features model and our new visual dictionary. At last, we normalize the final descriptors for videos and then recognize the actions using SVM. Experimental results show that our scheme achieves more accurate for action recognition compared with several state-of-the-art methods.