Article Preview
TopIntroduction
In recent years, there has been an exponential increase in the number of the videos recorded and shared over the internet. Consequently, the demand of analyzing and understanding videos automatically has also increased. Object tracking is a hard problem as many different and varying circumstances need to be included in one algorithm (Yilmaz et al., 2006). There is a plethora of work done in tracking of rigid objects like vehicles, ball, etc., under varying conditions but tracking of non-rigid objects is still a challenging task (Jalal and Singh, 2012). Fast tracking of deformable objects is necessary to handle flexible objects like rope, cloth, paper, human cell, sponge, and living creature. The tracking of deformable objects is useful in medical imaging, robotics, motion-based recognition, video indexing, etc. Tracking of such kind of objects can help us to train robot to clean the room, make us breakfast or help us in surgical operation. In spite of having broad range of applications, tracking large number of deformable objects simultaneously is very challenging, as target identity may be lost by the tracker (Chu and Smuelders, 2010).
The tracking of multiple moving objects has attracted large number of researchers to do research in this area of computer vision because of it large range of applications, to improve the quality of life for physical therapy patients and disabled people, to obtain automatic annotation of videos, to generate object based summaries and in traffic management to analyze flow and to detect accidents (Chu and Smuelders, 2010).
Various approaches for object tracking have been proposed. These mainly differ from each other based on the way they approach the following questions: Which representation of object is suitable for tracking? Which features should be used? How should the motion of the object be modeled? The answers to these questions depend on the circumstances and factors in which the tracking is performed. A large number of tracking methods have been proposed which attempt to answer these questions for a variety of context (Smeulders et al., 2014).
Due to the high dimensionality of the state spaces of deformable objects, perceiving deformable objects is much more difficult than perceiving rigid objects. Often, self-occlusions make it impossible to infer the full states of deformable objects from a single view (Schulman et al., 2013). In addition, many deformable objects of interest lack distinguishable key-points. While having many advantages (e.g., improving performance over time, online updating of the object and the background model, ability to derive a discriminative object detector at any time, etc.), the main challenge is to avoid drifting still being adaptive to changes in the object’s appearance. Sometimes, drifting problem also occurs due to improper annotation of object (Zhang et al., 2014).
In this paper, we present a novel detection method to detect the objects in the first frame of the video without using any background frame and tracking method that tracks the detected objects by exploiting spatial and temporal context. On the basis of this context model, the new location of the object is identified in the successive frames of the video. The rest of the paper is structured as follows. Section 2 discusses the previous work related to the tracking of deformable objects. Section 3 and 4 describes the main method of detection and tracking of object. In Section 5 results of our experiments are presented and section 6 concludes the paper.