Spatio-temporal continuous wavelet transforms for motion-based segmentation in real image sequences

The purpose of this paper is to develop a motion-based segmentation for digital image sequences that is based on continuous wavelet transform. Continuous wavelet transform allows estimating the motion parameters on all the moving discontinuities, edges and boundaries in the image sequence. The important fact in our case is that this technique provides all the information of motion parameter estimates and edge locations at once without going back and forth refining the segmentation and the motion parameter estimation. Also, this is achieved without involving any point/block corresponding techniques in our algorithm. The edges and the motion parameter estimates are calculated locally on small windows or pixels in the image planes by maximizing the square of the modulus of the wavelet transform. A clustering procedure allows separating all the detected edges into clusters of homogeneous motion. Building a ridge-skeleton on the reconstructed edges in each cluster provides the ultimate motion-based segments or partition. The algorithm was simulated using real traffic image sequences acquired by a mobile camera and proved to be accurate and robust.