Image Decomposition and Tracking with Gabor Wavelets

This paper explores the use of the Gabor wavelet representation of an image for robust object tracking in robot vision (object grasping, gesture recognition and face tracking in human-robot interaction). For image decomposition we developed a fast non-iterative transform algorithm, in which the original image is processed with a 2D Gabor wavelet filter bank. We used the positions of the local extrema (low-level object features) in the filter responses to calculate a set of Gabor wavelet coefficients. That guarantees on the one hand a low redundancy of the resulting representation and on the other hand an automatic detection of the significant features in the image. To prove the quality of the object representation the image can be reconstructed from the descriptor. It was shown that such descriptors can be used for real-time scale invariant object tracking by gradient following or template adaptation. Tracking control was realized by a scale factor discriminator control technique and convolution based 2D cross-correlation. Our experiments revealed that this method enables real-time object or face tracking in a wide scale range and provides robustness against high frequency camera vibrations, from which cameras on mobile vehicles suffer.