Abstract
The surveillance cameras have been largely deployed to record and track moving objects to prevent theft, illegal activities or whatever so as to secure the safety in public or private places prevalently. Tools for analyzing those surveillance videos should thereby be necessary and important for evaluating the behavioral patterns of humans identified as potential security risk. In this thesis, we focus on tracking multiple people simultaneously when they are interacting with each other, and then recognizing their activities as well. A novel motion-based approach, rather than the image-based, to effectively recognize the moving activities is proposed for the near-field visual surveillance. Key actions are extracted and characterized by motion directions of several portions (limbs, torso) of human body. Some activities, such as pushing, punching, kicking, etc. are strongly relative to these key actions associated with blobs of human limbs. We give some rules for identifying these key actions. After key actions on identified human blobs for two-person interactions labeled, we can recognize what activity they are doing by means of simple principle: The activity with the highest density in a fixed-time period will be the result. Finally, a SVM–based training procedure is employed to further refine the decision.