Tracking objects using Gabor filters Mark de Greef, Sjoerd Kerkstra, Roeland Weve Supervisor: Theo Gevers Overview Introduction Color spaces Feature selection and mean shift Demonstration Conclusion Tracking objects using Gabor filters 2 Introduction Goal: investigate the use of texture in tracking. Trackers often use color features (like RGB) Color features are not sufficient for tracking in some cases We use texture information as a basis for tracking Our tracker is based on the on-line feature selection framework. Tracking objects using Gabor filters 3 Color spaces RGB rgb HSV Intensity Tracking objects using Gabor filters 4 Gabor filters Captures localized frequency information Biologically motivated Tracking objects using Gabor filters 5 1D Gabor * = Tracking objects using Gabor filters 6 2D Log-Gabor * = Tracking objects using Gabor filters 7 Inverse Fourier transform IFFT -> Tracking objects using Gabor filters 8 Feature selection 1D histograms Background/object separation Tracking objects using Gabor filters 9 Mean shift tracking Probability density of target model and target candidate Try to minimize distance between probability densities Tracking objects using Gabor filters 10 Demonstration Tracking objects using Gabor filters 11 Tracking without log-Gabor 12 Tracking without log-Gabor 13 Tracking a grid ball 14 Tracking using log-Gabor Tracking objects using Gabor filters 15 Tracking a leaf Tracking objects using Gabor filters 16 Tracking without log-Gabor Tracking objects using Gabor filters 17 Tracking a leaf 18 Tracking with logGabor Tracking objects using Gabor filters 19 Tracking a soccer ball Tracking objects using Gabor filters 20 Tracking without log-Gabor Tracking objects using Gabor filters 21 Tracking a soccer ball Tracking objects using Gabor filters 22 Tracking with log-Gabor Tracking objects using Gabor filters 23 Problems High wavelength log-Gabor filters A frame in the soccer game movie Tracking objects using Gabor filters 24 Problems High wavelength filters pollute feature selection Solved by limiting wavelength of log-Gabor filter to 2*min(h,w) 25 Conclusion Log-Gabor filters make tracking possible in situations where color features do not Problems with large wavelengths can be avoided Addition of log-Gabor features to color features does not degrade tracking performance Tracking objects using Gabor filters 26 Questions? Tracking objects using Gabor filters 27