Optimization-driven machine-oriented motion measurement

Motion measurement is one of the most recent disciplines in the area of time-varying imagery which is one of the segments of the wider area of computer vision. In this area, because of many limiting factors in the data collection and the formulation of the motion problem, the two major trends in motion measurement, namely, the correspondence-based and differentiation-based techniques are regarded as fundamentally different and often generate different solutions starting with the same image data.

In this work, we propose a motion measurement model which unifies all these techniques, in particular the correspondence and 2- and 1- dimensional differential techniques. This optimization-driven machine-oriented motion measurement model regards these techniques as an optimization process where the image intensity data of a time-varying image is transformed into a 2-dimensional velocity field through the integration of a forward path representing the measurement extraction process from the pictorial data and a feedback path representing an additional set of topological and scene constraints imposed upon the input and/or output of this system.

This minimization process entails three major tasks: (1) identifying a set of tools to extract the amount of motion information that can be measured through the forward path, (2) identifying a set of additional constraints to supplement the latter information because the full velocity field cannot be recovered without utilizing the feedback path, and (3) identifying in which proportion the two former pieces of information should be mixed in order to recover the true velocity field. The core of this work is the identification of these three factors.

Using this newly introduced model, the motion problem is reformulated in a more general form than what is available in the literature for the correspondence and 1- and 2-dimensional differential approaches to motion measurement. In general, this formulation generates better and more efficient solutions by identifying potential motion clues and the amount of motion information that they carry. For instance, the velocity field measurement problem along simple closed contours can be formulated to generate a motion solution which is direct (as opposed to iterative) and computationally faster.