Visual Tracking System for Target Representation and Localization

Visual object tracking is the process in which objects in video sequences are tracked. The transformations the target object in motion undergoes like rotation, scale, change in view angle, illumination variations, partial and full occlusion in sequence of video frames have paved way for various representation methods and tracking algorithms. Here in this paper we propose a system for visual object tracking, which offers a representation of the targeted object depending on its appearance that is based on local steering kernel descriptor (LSK) as well as color histogram data. The input for the system is taken as the previous video frame as well as stored instance of object which is targeted. By searching the frame region we try to locate the target object in present frame which best resembles the target object input. The object model stores the instances of the target object and it gets updated as the view of the target object changes over time. The surrounding background as well as color histogram similarity in object which is decoded is utilized for the purpose of background subtraction. Given that the transformations in target object between two consecutive frames are rather small, the implemented system has been proven to be successful in tracking more than one target objects in a video and it also handles the case of full occlusion. We are keeping track of the LSK matrices of pervious frame and comparing it with the LSK matrices of the current frame to detect the target object and used Kalman filter to correct the predicted target object location.