Objects’ Segmentation And Tracking In Surveillance Videos

So far one of the most popular research topics in computer vision is visual surveillance indynamic scenes, especially for humans and vehicles. It has a large range application,including access control in special areas, human identification at a distance, crowd fluxstatistics&congestion analysis, detection of anomalous behaviors, and interactivesurveillance using multiple cameras, etc. Researchers have devised and developed manyautomatic surveillance systems according to their different applications and demands in recentyears, which has drawn a lot of attention of many researchers due to its broad applicability indifferent fields. In this thesis, the proposed algorithm may meet these demands. Theprocedure is divided into two parts. The first part is to segment the images into composedelements. And second part is to track the motion elements of the images.The first part of the work is to subtract the background from the images of the videosequences. We use Mixture of Gaussians Algorithm (MGA), which is a recursive, parametricmethod. It can track part of pixels following multiple Gaussian distributions simultaneously.This method keeps the density function for each pixel. Thus, it is capable of handlingmultimodal background distributions. On the other hand, since MGA is parametric, the modelparameters can be adaptively updated without keeping a large buffer of video frames. I haveapplied the algorithm on four different video sequences in different scenario including thecase of indoor and outdoor. The MGA allows us to identify foreground pixels in each newframe while updating the description of each pixel’s process.After the study of the performance of the MGA in different scenes and evaluate thedifferent parameters which give the best performance of the algorithm, the second part is toapply the object tracker to generate the trajectory of the tracked object over time by locatingits position in every frame of the video. To obtain much better results in tracking procedurewe add the prediction process by using the Kalman filter, which is useful to predict positionsof points in two dimensional images. We apply Kalman filter on these different videos andverify the contribution of the filter on accurate tracking.