Time-Varying Motion Estimation on a Sequence of Images

This work is concerned with the estimation of time-varying motion fields in a sequence of images. We consider only the ``apparent'' motion of any point in the scene, i.e. its projection over the image plane, rather than the three dimensional motion of points in the scene. Unlike many other approaches that use simply a motion model with constant velocity for the pixels in the image, a second-order time-varying motion model is considered. We believe that this approach is much more suited to describe the motion of real-world objects, as their motion is obeying the fundamental laws of classical mechanics. Certainly the model assumes that objects in the scene are subject to a constant acceleration within reasonable periods of time. If this hypothesis remains true, this can provide a way to ease the tracking and identification of moving objects in a scene, making more robust any subsequent segmentation task (assuming that one can correctly estimate the parameter of the motion model, and identify the interval of time during which the object remains under a constant acceleration). To bear with our hypothesis, we will show how this second order motion model outperforms the classical constant-velocity model for predicting frames of an image sequence. This could suggest that with very few parameters, the motion of portions of an image sequence can be correctly modelled with significant impact in a video compression scheme or in a frame rate conversion application.