Date of Award
CC BY-NC-ND 4.0
The aim of thesis is the real-time detection of moving and unconstrained surveillance environments monitored with static cameras. This is achieved based on the results provided by background subtraction. For this task, Gaussian Mixture Models (GMMs) and Kernel density estimation (KDE) are used. A thorough review of state-of-the-art formulations for the use of GMMs and KDE in the task of background subtraction reveals some further development opportunities, which are tackled in a novel GMM-based approach incorporating a variance controlling scheme. The proposed approach method is for parametric and non-parametric and gives us the better method for background subtraction, with more accuracy and easier parametrization of the models, for different environments. It also converges to more accurate models of the scenes. The detection of moving objects is achieved by using the results of background subtraction. For the detection of new static objects, two background models, learning at different rates, are used. This allows for a multi-class pixel classification, which follows the temporality of the changes detected by means of background subtraction. In a first approach, the subtraction of background models is done for parametric model and their results are shown. The second approach is for non-parametric models, where background subtraction is done using KDE non-parametric model. Furthermore, we have done some video engineering, where the background subtraction algorithm was employed so that, the background from one video and the foreground from another video are merged to form a new video. By doing this way, we can also do more complex video engineering with multiple videos. Finally, the results provided by region analysis can be used to improve the quality of the background models, therefore, considerably improving the detection results.
Kaur, Shanpreet, "Background Subtraction in Video Surveillance" (2017). Electronic Theses and Dissertations. 5944.