Demons Based Tracking for Non-Rigid Transformed Region of Interest Abhinav Kumar , B Madhusudan Rao, Rajesh Ghole, Amol Patil 1 , Nilesh Ghatpande CoE-Image Processing, iGATEPatni Pune, India 1 amol.vpatil@igatepatni.com Abstract— Tracking of objects or image region undergoing non- rigid transformation is a challenging and central problem of computer vision. It gets further complicated when the object/region to be tracked is similar to other nearby objects/regions or background. This paper presents a Region Of Interest(ROI) tracking approach based on Maxwell’s demon based image registration algorithm. This approach doesn’t require features of an object/region to be extracted, but rather works only on the pixel intensities. This enables it to be suitable for tracking object/region undergoing non-rigid transformations and having little contrast with the background. The extensibility of the current approach to more complex problems like multiple ROI tracking and to handle almost any arbitrary changes in ROI is evident. We demonstrate the proposed non-rigid ROI tracking algorithm using endoscopy video data which is one of the potential applications of proposed algorithm. Keywords- Computer Vision, Demons Algorithm, Object Tracking, ROI Tracking I. INTRODUCTION Tracking of objects in a video is a well researched topic in the domain of computer vision. Some of the well known algorithms in this domain are, mean shift tracking [1], CAM shift tracking [2], optical flow based tracking [3] etc. These algorithms are based on extraction of features related to object of interest in reference image and perform matching using deterministic or stochastic criteria. For example, one of the object features is edges with high gradient value. Thus, the description of the object of interest is limited to extracted features. Even though these techniques can achieve tracking, they are likely to suffer from non-uniqueness of tracked objects resulting in erroneous behavior. This limits the utilization of the above stated algorithms. Established methods for object tracking are based on similarity measures (simplest being SSD or SAD) and object descriptors. Hence these approaches are feasible if the object of interest can be well defined or described. In other words, these approaches are suitable when the object (with its background) undergoes affine and simple non-rigid transformations. Under similar conditions, complexity of the tracking problem increases drastically if an object undergoes arbitrary non-rigid transformation [4]. An object by definition has identifiable features with its background such that it can be well described. However, if we consider a more generic case where, a region is marked for tracking without any assumption about its contents, then description of such ROI may not be achieved using a simple model. Moreover, a typical ROI can contain none to many objects which makes tracking of ROI undergoing arbitrary non- rigid transformations a challenging problem. The problem of tracking non-rigid transformed ROI can be defined as follows: A ROI is defined in a reference frame and the objective of tracking algorithm is to track each pixel of the ROI in all consecutive frames. Such a requirement may come up in several different applications. For these type of problems, there have been different approaches primarily based on matching of extracted features [4] [5]. In this paper we present an algorithm for automated tracking of user defined ROI. The input data considered are endoscopic videos from actual endoscopic procedures [6]. Following assumptions are made in the present implementation of the proposed approach. 1. Algorithm is applied on only one color channel from the color video input. 2. The main source of deformation of ROI is camera movement and change in its angle. 3. The scene being recorded is not undergoing any rapid movement or change. This ensures smooth transition between consecutive frames of video. The authors propose an algorithm for this problem by modifying an image registration based approach. This approach has been followed because the specified registration technique provides ability to track each pixel in the form of displacement field. The registration algorithm used here is based on Maxwell’s Demons [7]. The next section provides a brief overview of the same. In section III, the algorithmic modifications required for tracking an ROI are presented. Results of the numerical implementation along with details of convergence properties of the algorithm are presented in section IV. Finally the conclusions, discussion and future works are covered in the last section. II. DEMONS ALGORITHM FOR IMAGE REGISTRATION Maxwell’s demons were hypothetical entities introduced by Maxwell in 19 th century as part of a thought experiment in the area of thermodynamics and statistical mechanics. These 978-1-4577-0255-6/11/$26.00 ©2011 IEEE 321 TENCON 2011