Recognition

Recognition of motion using single view constraints was reported earlier [11]. Single view constraints can be used to index into motions of a person performing a sitting movement, for instance. The same philosophy can be used for recognition using our constraints, which need less points.

It is also possible to recognize objects undergoing non-rigid motion using these constraints, as long as parts have uniform velocity or uniform acceleration. Suppose we can track $N$ sets of $4$ points on the deforming body in all the views. From $8$ frames of each view and each set of points, a set of $\beta$'s can be computed. The $\beta$'s for a set of points computed from any view will satisfy the constraints expressed in Equation 21 for that set of points, for all views. This should be true for all $N$ sets of points, for the body to be the same in all the views. This recognition strategy was tested on views of the exploding tea pot scene. A sets of points on the tea pot were tracked across the explosion in all the views. It was found that the $\beta$ values computed for each set of points from one view were valid for the same set of points in all views. The coefficients computed were highly similar for a specific motion and different for different motion. These coefficients may be used for gait recognition or similar applications.