Model free object tracker

Abstract

In recent times, a lot of focus has been diverted to model free object tracking instead of model based. Model free tracking requires no prior model of object of interest. The object is represented by its location and size which is called initial annotation. The task of tracker is to build a correspondence of this object in all consecutive frames. The tracker follows this object and tries not to loose the track in whatever difficult situations like clutter, object hiding or pose changes. In this thesis, a novel and efficient object detection and tracking method is proposed keeping in consideration the real time requirements. The combination of two features Local Binary Pattern (LBP) and Histogram of Oriented Gradients (HOG) is applied for object localization. Both of these feature complement each other and their combination when fed to a linear Support Vector Machine (SVM) classifier would lead us to a robust learning measure. Our main focus, in this study, is on decreasing the time required to process each frame. The target object is represented by a parity based linear combination of HOG and LBP feature. A robust multi-modal learning mechanism is used for a more discriminative hypothesis generation. We have evaluated our approach with the widely available visual tracker benchmarking videos and have found our technique to be working efficiently.