|M.Sc Student||Hason Eshel|
|Subject||Tracking Unmanned Aerial Vehicles Using Active Contours and|
|Department||Department of Electrical Engineering||Supervisor||Mr. Allen Robert Tannenbaum|
|Full Thesis text|
This paper presents a novel algorithm for tracking UAVs (Unmanned Aerial Vehicles) in video sequences for real-time military applications. The scenario dealt by the algorithm is a single flying UAV photographed (in standard video rate, about 25-30 frames per second) against a stationary background. Although the UAV’s velocity may be high, its dynamics are fairly slow compared to the video rate. That means that the changes in the UAV’s shape and direction of movement between consequent image frames are moderate. In order to achieve the required ability of real-time implementation, the emphasis of the algorithm is on low complexity and fast target acquisition rather than high accuracy. The algorithm is based on a combination of an active contour method in level set formulation as a segmentation algorithm, Kalman filter as a tracking algorithm and motion detection for acquisition.
First, we survey and discuss the advantages and weaknesses of two classic segmentation methods: Bayesian segmentation (Maximum Likelihood/Maximum Posterior probability estimation) with non-linear, edge preserving filtering; and active contours (mainly in level set formulation). We also survey and discuss some classic tracking methods such as the Kalman filter and the basic particle filter.
We then present the general flow of the suggested algorithm and specify in detail the optimized system model for the Kalman filter tracking mechanism (a random acceleration model in both axes, both for target location and for target size, decomposed into four independent one dimensional random acceleration models), the active contour segmentation mechanism (based on minimization of luminance variances of both inside and outside the contour), and the motion detection acquisition mechanism (based on detection of luminance changing pixels in sequential image frames).
Finally, we present the results of the algorithm on stimuli of a light airplane maneuver video sequence, and confirm that the algorithm’s performance is adequate. We discuss the strengths and weaknesses of the algorithm and set the grounds for future research in the subject.