Unsupervised motion segmentation of image sequences using adaptive filtering

In this paper we present a new method for feature extraction in unsupervised motion segmentation of image sequences. This method is based on the multichannel filtering of the input image sequence using adapted Gabor filters. We use the histogram of quantised two-dimensional parameter space resulting from a three dimensional orientation analysis to select the main Gabor filter parameters, namely azimuthal and elevational angle. The three-dimensional orientation analysis consists essentially of the eigenvalue and eigenvector computations of inertia tensors in the three-dimensional frequency space. The feature images are obtained from the comlex valued filtered sequences using a simple magnitude computation and are subsequently evaluated employing a multichannel segmentation algorithm. The performance of the algorithm ids demonstrated using two segmentation examples that are based on artifical as well as real image sequences