Object detection and tracking from surveillance videos

Wide use of video surveillance systems calls for powerful tools to extract information from video data. In this dissertation, object detection and tracking algorithms are the focus of study. Three different object detection and two different tracking algorithms, which have gained their popularity in computer vision research community, are investigated in this dissertation. The three object detection algorithms investigated in this dissertation are background subtraction with adaptive Gaussian mixture model, Histogram of Oriented Gradients (HOG) detector and Deformable Part Model (DPM) detector. Background subtraction with Gaussian mixture model can detect moving objects fast and accurately in static environment. HOG detector and DPM detector can discriminate objects over background if trained before. For object tracking, the classical algorithm, Kalman filter, is studied. The Tracking-Learning-Detector (TLD) algorithm is also studied, which is a powerful tool for long-term detection and tracking. In this dissertation, the above algorithms are evaluated on two standard benchmark datasets, i.e. TUD-Stadmitte and PETS-2009 S2/Ll. The performance of these algorithms are reported and discussed. Master of Science (Signal Processing)