Video Surveillacne System For Marine Scenes

In this paper, the network video surveillance technology for marine scenes is discussed. We access the video streams of marine scenes such as fairway and ports on Internet and do real-time motion detection on user-defined surveillance regions to discover the prohibited moving objects and alarm according the scene sensitivity.We adopt the Mixture of Gaussians approach to model the scene for motion detection. A set of Gaussians are created for each pixel location and the current pixel is classified as background or foreground based on its probabilities sum accepted by Gaussian models which are updated with current pixel value. A mean-filtering denoising strategy is used limited on initially detected foreground pixels and works accurately and fast. In order to track multiple objects, we construct a model including color, shape and motion for each detected object area and create a distance matrix and matching matrix between newly-detected-objects and trajectories. If an unmatched object overlaps with more than one trajectory, it can be consider as a merger of two occluded trajectories. The model similarity between objects and components of a merged-trajectory is computed to test whether two merged objects split into individuals. This merge-split approach has considerable effect on handling objects occlusion.This paper proposes a Gaussian model for the foreground area in every frame to detect the abrupt camera motion. In addition, color histogram of the current frame is calculated at the interval of 5 seconds to percept the weather and illumination change which influence the accuracy of surveillance system so as to eliminate false alarm.The video surveillance system for marine scenes in this paper supports real-time monitoring as many as 25 channels of video streams. Different sensitivities are designed to apply in different scenes robustly. What's more, the system can handle objects occlusion and provide the objects statistical function if necessary.