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.