AN APPROACH TO TRACK MULTIPLE OBJECTS

Abstract

Object Tracking is one of the most important processes in the area of computer vision. Its objective is to locate a moving object or several objects in video sequences, using a single camera or multiple cameras. A tracking algorithm is used, whose purpose is to analyze the video frames in isolation and predict the location of moving targets in the video frames. So, we can define tracking as the task of localizing an object of interest from a video scene by means of segmentation and afterwards keeping track of its motion, orientation, occlusion etc. in the successive frames. This is done to extract useful information from the video sequences which can be used for various purposes, for example, traffic surveillance, crowd monitoring, robot navigation and many such applications. There are three key steps in object tracking: detection of interested moving objects, tracking of such objects from frame to frame, and analysis of object being tracked to recognize their behavior. A myriad of algorithms has been developed to implement these subtasks but each one has its own strengths and weaknesses. Extensive research has been carried out in this field over the last years to find optimal tracking systems for specific applications. Many approaches to track objects have been proposed in the literature, however, they are not completely accurate for all kinds of scenarios and just provide good results when a certain number of assumptions are verified. This is because sufficient information about object dynamics and features is generally not available beforehand in real time applications of object tracking. This causes errors in tracking and reduces the reliability of robustness of the tracking system. Different authors have tried different methods for object tracking like particle filters, background-foreground segmentation approaches and so on. This thesis aims to provide a graph cut based technique for tracking multiple objects in the presence of occlusions to get optimal results in terms of accuracy, robustness and reliability as compared to existing techniques.