Implementation of Background Modelling and Evolving Fuzzy Rule – based Classifier for Real-Time Novelty Detection and Landmark Recognition

The important role of surveillance and control systems in maintaining human safety in nowadays life creates a vast field for designing algorithms to fulfill security. This thesis demonstrates the procedures of implementing background modeling and evolving fuzzy rule-based classifier (eClass) for real – time novelty detection and object tracking. Initially the ways which security systems perceive the information obtained from an outlook are described along with the basic design of a system for tracking the object. Kernel density estimation method is illustrated and performed as an advanced online approach which models the background and foreground of the scene in order to enable the system to be sensitive to important interactions happen in a location by evaluating the probability of information based on the data history. Illustrating significant values of machines performing autonomously and unsupervised is another aim of this project. In certain circumstances like unknown environments, human interaction is risky; therefore, developing algorithms to facilitate machines to execute certain tasks without having prior knowledge about the environment is accomplished in this report. Machines are equipped by algorithms with fuzzy logic characteristics in order to think more similar to human mind and learn by themselves. This thesis depicts a novel approach to novelty detection and object tracking with evolving property which can improve and update itself gradually by learning without supervision. Therefore eClass algorithm is explained in this report as a method to image segmentation and self-localization which consequently paves the way for object tracking. Tracking the object using eClass as a fast and recursive method is compared with the kernel density estimation and the advantage and drawbacks of each are presented. Ultimately, Kalman filter and evolving extended Takagi-Sugeno models are demonstrated in order to reveal the use of prediction module in surveillance systems and are compared with together to present their efficiency and accuracy in terms of correlation between the obtained predictions and real values.