Tracking Using Intensity Gradients and Particle FilteringJason ReneauDepartment of Electrical and Computer EngineeringClemson UniversityClemson, SC 29630AbstractAccurate tracking of an object in a videosequence undergoing dynamic motion canbe a challenging task from a non-movingplatform. When the object to be tracked andthe tracking platform are moving it becomesmore difficult to successfully track the object.Background clutter, lighting variations andhuman operation of the camera canincrease the tracking problem difficulty to thepoint where little usable video may beobtained. The objective of this project is toexamine the use of an intensity gradienttracker in combination with a particle filter toachieve accurate tracking of an aerial targetfrom a moving platform in an existing videosequence. A secondary goal of this projectis to evaluate possibility of using thistechnique in a real-time system to control acamera to autonomously track the objectwithout human interaction. The results ofthe application of an intensity gradient andparticle filter tracking approach arepresented and analyzed.1. IntroductionIt is desirable to track aerial targets withvideo cameras in order to evaluate theiraerodynamic qualities. Accurate trackingcan be difficult to achieve when the targetbeing tracked and platform tracking theobject are experiencing dynamic motion.Other factors that complicate the trackingproblem are background clutter in the scene,lighting variations and human cameraoperators. It is not always possible to trackthe object from a second aircraft with ahuman camera operator due to safety orother concerns. The objective of this projectis to examine the possibility of using animage gradient tracker in conjunction with aparticle filter to achieve accurate tracking ofan aerial target in a video sequence. Thevideo sequence being examined here wastaken from a second aircraft by a humanoperator. The results of the proposedtracking technique will be used to drawsome conclusions as to the feasibility ofusing the technique in a real-time trackingsystem. The object to be tracked is modeled asa rigid two dimensional making it possible touse an intensity gradient tracker. Trackingobjects modeled as rigid two dimensionalobjects has been successfully demonstratedin real-time systems [2]. An algorithm usingthe intensity gradient and modeling the headas an ellipse has been used to track peoplethrough a video sequence [2]. The simplemodel, straightforward implementation andreal-time results achieved with the gradienttracker were the reasons it was selected forthe aerial tracking problem posed in thispaper. The intensity gradient tracker doeshave some drawbacks. In order for it toexecute quickly enough for real-timeapplications a local search of the intensitygradient is performed. Dynamic motion cancause the tracker to lose track of the objectif it moves out of the search area. As aresult a constant velocity motion model hasbeen used to provide a more reliabletracking [2]. Rather than use the motionmodel described in [2] a particle filterapproach is proposed to predict the state ofthe object being tracked. The particle filteris able to track dynamic or agile motionbecause it applies dynamic models withobservations to propagate a random setthrough time [5]. The particle filtering orCondensation algorithm has demonstratedthe ability to track objects experiencingdynamic motion in cluttered environments innear real-time [5]. The video sequenceexamined in this project the object beingtracked experiences dynamic motion due tothe object itself and movement of thecamera platform. There is very little clutter inthe video to confuse the tracker. 2. MethodThe first image of the video sequence issearched globally for the location thatmaximizes the intensity gradient for themodel of the object to be tracked. The initiallocation and size of the object in the firstimage is used to initialize a local intensitygradient search of the second image. Aparticle filter algorithm is implemented toestimate the state of the tracked objectbased on the measurement of the intensitygradient. The state estimated by the particlefilter is used to update the local gradientsearch in the next image. Subsequentframes of the video are searched by theintensity gradient tracker based on theprobabilistic model estimated by the particlefilter until all of the frames of the videosequence have been tracked [3]. 2.1 Intensity GradientThe intensity gradient tracker uses a 2Drectangle with a fixed aspect ratio to modelthe object to be tracked [2]. The rectangle’sstate is characterized by three parameters s= (x, y, k), where (x,y) represent the centerof the rectangle and k the length of therectangle. A fixed ratio is used to determinethe size of the rectangle [2]. The trackerperforms a search of the image thatmaximizes the normalized sum of thegradient magnitude around the rectangle:NiiSsgNs11maxarg (1)gi is the gradient at pixel i in the perimeterand N is the total number of pixels in therectangle [2]. Both the global and localsearch intensity gradient calculations usethe summation of equation (1) to construct agradient map of the image and locate the(x,y) track point. The global intensity gradient searchesfor the initial position and for the k value thatmaximizes the normalized sum of theintensity gradient over the image. Theimplementation used here assumes that theborder region will not contain the object tobe tracked. The assumption was made tosimplify the code for testing but would not bemade in a practical system.The local intensity gradient search isbased on the state estimated by the particlefilter. A window of the image around theestimated (x,y) position