Book Description

The growth in the number of surveillance cameras deployed and the progress in digital technologies in recent years have steered the video surveillance market towards the usage of computer systems to automatically analyse video feeds in a collaborative and distributive fashion. The semantic analysis and interpretation of video surveillance data through signal and image processing techniques is called Video Analytics (VA). In this thesis new video analytics methods are presented that are shown to be effective and efficient when compared to existing methods. A novel adaptive template matching algorithm for robust target tracking based on a modifed Sum of Absolute Differences (SAD) called Sum of Weighted Absolute Differences (SWAD) is developed. A Gaussian weighting kernel is employed to reduce the effects of partial occlusion, while the target template is updated using an Infinite Impulse Response (IIR) Filter. Experimental results demonstrate that the SWAD-based tracker outperforms conventional SAD in terms of efficiency and accuracy, and its performance is comparable to more complex trackers. Moreover, a novel technique for complete occlusion handling in the context of such a SWAD-based tracker is presented that is shown to preserve the template and recover the target after complete occlusion. A DSP embedded implementation of the SWAD-based tracker is then described, showing that such an algorithm is ideal for real-time implementations on devices with low computational capabilities, as in the case of xed-point embedded DSP platforms. When colour is selected as target feature to track, the mean shift (MS) tracker can be used. Although it has been shown to be fast, e ective and robust in many scenarios, it fails in case of severe and complete occlusion or fast moving targets. A new improved MS tracker is presented which incorporates a failure recovery strategy. The improved MS is simple and fast, and experimental results show that it can effectively recover a target after complete occlusion or loss, to successfully track target in complex scenarios, such as crowd scenes. Although many methods have been proposed in the literature to detect abandoned and removed objects, they are not really designed to be able to trigger alerts within a time interval defined by the user. It is actually the background model updating procedure that dictates when the alerts are triggered. A novel algorithm for abandoned and removed object detection in real-time is presented. A detection time can be directly specified and the background is "healed" only after a new event has been detected. Moreover the actual detection time and the background model updating rate are computed in an adaptive way with respect to the algorithm frame processing rate, so that even on different machines the detection time is generally the same. This is in contrast with other algorithms, where either the frame rate or the background updating rate is considered to be fixed. The algorithm is employed in the context of a reactive smart surveillance system, which notified the occurrence of events of interest to registered users, within seconds, through SMS alerts. In the context of multi-camera systems, spatio-temporal information extracted from a set of semantically clustered cameras can be fused together and exploited, to achieve a better understanding of the environment surrounding the cameras and monitor areas wider than a single camera FOV. A highly flexible decentralised system software architecture is presented, for decentralised multi-view target tracking, where synchronisation constraints among processes can be relaxed. The improved MS tracker is extended to a collaborative multi-camera environment, wherein algorithm parameters are set automatically in separate views, upon colour characteristics of the target. Such a decentralised multi-camera tracking system does not rely on camera positional information to initialise the trackers or handle camera hand-o events. Tracking in separate camera views is performed solely on the visible characteristics of the target, reducing the system setup phase to the minimum. Such a system can automatically select from a set of views, the one that gives the best visualisation of the target. Moreover, camera overlapping information can be exploited to overcome target occlusion.