Performance evaluation of human activity recognition in video sequences

Abstract

In this thesis, we address human behavior recognition, as one of the important topics in computer vision. It finds applications in many areas such as surveillance, military installations, and sports. The problem becomes more challenging, due to the huge intra-class variation, background clutter, occlusions, illumination changes and noise. Human behavior recognition typically requires standard preprocessing steps such as motion compensation, background modeling. The errors of the motion compensation step and background modeling increase the mis-detections. We use JBFM as our background model and optic flow values to compute the motion. We propose two different spatio-temporal feature descriptors, SOF and DTF, which combine both computed motion and appearance based features. We use SVM to recognize human actions, by using different evaluation protocols (test cases). We perform several experiments and compare over a diverse set of challenging videos to address the problem, human behavior recognition by simplifying into three tasks. They are, human action recognition in stationary background, human action recognition in dynamic background, and abnormal activity recognition. Our Experimental results show that the selected framework outperforms state-of-the-art methods in many cases in terms of both recognition rate and computational complexity.