Abstract
Recently, many active vision techniques for object recognition are reported in the literatures. In the technique, the shape of an object is reconstucted using the motion parameters of the camera and the image optical flows of the object points. However, obviously, the motion parameters must be measured exactly for the 3-D shape reconstruction. In usual cases, it is difficult to know such exact motion parameters by using the sensors attached to the camera. But, the image pixels taken from the camera have much more accurate capability to locate the point positions than those sensor's outputs.
In this paper, we propose a new method of active vision which can recognize the 3-D shape of an object without knowing exact camera motion parameters. The motion parameters are calculated from the optical flows and the depth of object points whose 3-D shape is already known. And, using this calculated motion parameters and the optical flows, the 3-D position of unknown points are reconstructed. Repeating these two phases, taking the image of a known object, 3-D shape of unknown objects are reconstructed as camera moves. In this method, only the sequence of images are used to determine the camera motion and the 3-D point positions.
In this method, the effects of quantization errors, which cannot be avoided in digital image analysis, will be overcome by two approaches. As for camera motion parameters, the errors are compensated by using a large number of points to calculate the position and the velocity of the camera. Then, the Kalman filtering method is applied to the sequence of images to reduce the 3-D position errors of each unknown points.
Results of experimental simulations are shown to confirm the effectiveness of this proposed method.
