Recognition Of Suspicious Activity In Video Surveillance

Abstract:Currently, recognition of suspicious activity in video surveillance has become an important issue of computer vision. This thesis focuses on several related problems which are suspicious activity preliminary detection, segmentation of the activity, human action recognition and suspicious activity recognition. Our contribution is concluded in the following aspects.(1) A new method based on saliency-based visual attention model is proposed to detect suspicious activity. Due to the high computational complexity and the difficulty in the common use of the existing algorithms, the proposed method considers all the activities in the video as a whole and detects suspicious activity using the visual attention model based on spatial-temporal saliency rather than analysis each activity respectively as the existing algorithms.(2)The method of suspicious activity decomposition based on probabilistic principle component analysis (PPCA) is presented in this thesis. We use PPCA to model the characteristics of human silhouette. As a result, different action characteristics have different distribution. According to the different distribution, we can find the decomposition point. The method does not require complex model of the activity and can decomposite suspicious activity quickly and efficiently.(3) For the actions decomposted from the suspicious activity, this paper introduces action recognition based template matching. Zernike moment is used as the similarity measure which effectively represents the translation, scale and rotation invariants of the characteristics. Compared to the state model method, our method avoids complex parameter setting and model training process, which can improve the performance of action recognition.(4) A novel method based on Case-Based Reasoning is proposed for suspicious activity recognition. The case of suspicious activity is represented by sub-behaviors included in the activity and take the time-associated characteristic of the sub-behaviors into consideration. Suspicious activity is identified by the matching algorithm. The method based on case-based reasoning is superior to the method based fixed model and rule because of the application of a dynamic database.Experiment results show that the method porposed in this thesis can efficitively recognize suspicious activity in video surveillance.