Automatic scene analysis and reconstruction from video

Abstract

For humans, vision is one of the most important and complicated senses. Since the first digital computers were created, various attempts have been conducted to attribute sight to computers. Vision would allow a computer to understand its environment and distinguish objects in its field of vision. The need for greater security in public places led to the instalment of numerous surveillance cameras. These cameras need to be constantly monitored by a human person. A computer capable of understanding its surroundings would be able to partially or totally replace the human factor. This thesis investigates the different methodologies used to process video sequences for the scope of human tracking. The execution of two major background modelling techniques, Running Average and Gaussian Mixture Model were compared. A background modelling technique outlines the foreground pixels from the background scene. A Shadow Removal algorithm using the Hue, Saturation and Value (HSV) colour was employed to remove shadow pixels that may hinder the overall tracking process. Morphology filtering techniques were also implemented to reduce any pixel misclassification. Finally the Correspondence-based Object Matching technique was implemented to find a correspondence between objects detected in two consecutive frames. Each moving object features are extracted to help in the drawing of statistical analysis regarding the image sequence. The average size, average speed and number of objects tracked per frame are presented in a graphical format. Moreover, the implemented system outputs a tracking and heat map video sequence illustrating the tracked objects and the most and least frequented areas in the scene respectively.