Abstract
Spatio-temporal interest points in video provide a compact and representative information for the video analysis such as motion recognition. Most of the previous spatio-temporal interest point detection methods used only local information and their detection results are not robust. Recently, a new interest point detector based on applying NNMF (Non-Negative Matrix Factorization) on a video sequence to obtain the spatial subspace matrix and temporal coefficient matrix was proposed and proved to significantly improve the accuracy of several motion recognition tasks. However,this NNMF-based interest point detector takes a considerable amount of execution time. In this thesis, we propose a PCA-based algorithm for spatio-temporal interest point detection and selection from a video sequence. The proposed algorithm is composed of three stages. At the first stage, we factorize the video sequence into spatial subspace matrix and the corresponding temporal coefficient matrix. Then we detect the interest points on the 2D spatial subspace eigen-images and the 1D temporal coefficient vectors separately. Finally, we apply a saliency measure to select representative interest points from all the pairings of the detected points. Experimental results on facial expression recognition show the motion classification based on the proposed interest point detector based on PCA-based matrix factorization can provide satisfactory accuracy and its computational speed is much faster than the NNMF-based interest point detector.