This thesis addresses the issues of applying advanced video analytics for surveillance applications. A video surveillance system can be defined as a technological tool that assists humans by providing an extended perception and capability of capturing interesting activities in the monitored scene. The prime components of video surveillance systems include moving object detection, object tracking, and anomaly detection. Moving object detection extracts the foreground silhouettes of moving objects. The object tracking component then applies the foreground information to create correspondences between tracks in the previous frame and objects in the current frame. The most challenging part of the system concerns the use of extracted scene information from the moving objects and object tracking for anomaly detection. The thesis proposes novel approaches for each of the main components above. They include: 1) an efficient foreground detection algorithm based on block-based detection and improved pixel-based Gaussian Mixture Model (GMM) refinement that can selectively update pixel information in each image region; 2) an adaptive object tracker that combines the merits of Kalman, mean-shift and particle filtering; 3) a feature clustering algorithm, which can automatically choose the optimal number of clusters in the training data for scene pattern classification; 4) a statistical scene modeller based on Bayesian theory and GMM, which combines object-based and local region-based information for enhanced anomaly detection. In addition, a layered feedback system architecture is proposed for using high- level detection results for improving low-level detection performance. Compared with common open-loop approaches, this increases the system reliability at the expense of using little extra computation. Moreover, considering the capability of real-time operation, robustness, and detection accuracy, which are key factors of video surveillance systems, appropriate trade-offs between complexity and detection performance are introduced in the relevant phases of the system, such as in moving object detection and in object tracking. The performance of the proposed system is evaluated with various video datasets. Both qualitative and quantitative measures are applied, for example visual comparison and precision-recall curves. The proposed moving object detection achieves an average of 52% and 38% improvement in terms of false positive detected pixels compared with a Gaussian Model (GM) and a GMM respectively. The object tracking component reduces the computation by 10% compared to a mean-shift filter while maintaining better tracking results. The proposed anomaly detection algorithm also outperforms previously proposed approaches. These results demonstrate the effectiveness of the proposed video surveillance system framework.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:555815 |
| Date | January 2011 |
| Creators | Li, Hao |
| Publisher | University of Bristol |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
Page generated in 0.0021 seconds