New Spatiotemporal Approach for the Estimation of the Motions of Multiple Images using Time-Frequency Representation

Estimating the motion characteristics within a time-varying scene and identifying the moving objects is an important task in many video applications. Most of the proposed motion analysis techniques are carried out directly in the pixel-intensity domain, applying typically the optical flow equation or correlation-matching in short sequences of frame, combined with spatial motion-smoothness constraints in order to obtain the actual motion field. How this has the unwanted effect of over-smoothing the motion discontinuities, across object edges. More recently, this problem is partially resolved within a robust estimation framework. The technique presented in this paper belongs to the class of spatiotemporal (3-D) spectral approaches, which exploit the information contained in long image sequences and estimate directly the true motion characteristics, regardless of the moving objects’ shape or dense motion discontinuities. However, the limitation of most spatiotemporal approaches is the assumption of time constant motion. Thus, they can be applied only for short time intervals. A few attempts have been made to overcome the limitation. The authors in address only the problem of a single global time-varying motion, using the high-order ambiguity function, to estimate the parameters of chirp signals. The work in handles multiple accelerated motions, exploiting the chirp-Fourier transform and a clustering algorithm. In another recent work, the instantaneous velocities are estimated using time-frequency representations (TFRs), with the restriction, however, of stationary background and small moving objects. Combining and mainly extending previous relevant ideas, this paper presents a “motion first” spatiotemporal methodology, for the estimation of multiple, occluded or transparent, time-varying motions and the extraction of the moving objects in the frequency domain.