A method for recovery of the 3‐d structure and motion of objects from low‐accuracy dynamic images

Abstract

AbstractWhen observing a time‐series of two‐dimensional images composed only of points on a moving object, a human can perceive the 3‐D structure essentially existing among the points, and the object itself can be recognized if a sufficient number of points exists. For such a stereokinetic visual function, numerous recovery methods have been proposed for 3‐D information from 2‐D images. However, past methods of recovery have a problem in that the accurate recovery is impossible unless the image data have a high accuracy. For example, the input data must have the accuracy of 10 decimal digits. On the other hand, the human seems to perceive the 3‐D structure of the object from the data with a relatively low accuracy by another means, such as a long‐term observation.This paper discusses a method of recovery where the 3‐D structure is recognized with as high an accuracy as possible from the data with a relatively low accuracy using a number of frames. Rigid bodies connected by joints are considered as the objects. When the information concerning the positions and velocities of several points on the object projected on a plane are provided, the depth information is recovered. By a computer simulation, it is seen that if approximately 10 or more frames are available, the output result with one order of magnitude lower accuracy than that of the input data can be obtained.