This paper presents a solution to the optimal camera movement for 3-D shape recognition.
Our solution method involves the following steps. We capture sequence of images from the moving camera and then recognize the 3-D shape of the objects by applying Kalman filtering. Each object is identified by a few points on the object, which we term the object points. We are able to estimate the accuracy of the object point location by means of the corresponding co-variance matrix.
Depending on this accuracy, we assign recognition priority to the objects. Unrecognized objects are assigned higher priority values compared to the objects which are partially recognized or completely recognized. Note that our optimal camera movement depends on the priority values. But the priority values change dynamically, and also new objects appear in the image view. That means the camera movement also has to be decided dynamically. The camera movement is viewed as the composition of translational and rotational movements. The angle made at an object point by the translational motion of the camera is defined as the disparity angle of the object point. The translational component of the camera should be such that the sum of the disparity angles weighted by the priority values of the object points must be maximum. However, if the translational component turns out to be too large, it becomes difficult to correspond the object points. The movement, therefore, should be restricted within some small value.
The rotational component, on the other hand, should be such that the objects must be retained within the image view for a long time. Because, to converge appropriately, Kalman filter requires a large number of images of the same object.
The method provided here is an efficient solution of dynamically determining the camera movements fulfilling all the above requirements.
抄録全体を表示