A predictive architecture is presented to react on sensor data in the case of high speed motion and low bandwidth sensor data. This concept is used for the vision based control of an industrial robot to track a contour at a speed of 1.6 m/s. The vision task can be performed very fast since only 2 rows of the image are analyzed. In this way an accuracy of 0.3 mm is reached in spite of uncertainties in robot's kinematic parameters. Vision and control work asynchronously so that even delay times are tolerable during sensing as long as the time-instant of the exposure is known.

Position-based visual servoing requires estimating and tracking the three dimensional position and orientation of a 3D target object from camera images. This paper describes a novel approach to the problem that consists of two steps. First, a set of spatial pose constraints is derived from image features, by means of which 3D object pose is calculated with an efficient model-fitting algorithm. Kalman-filtering is then used to estimate object velocity and acceleration. Compared to previous approaches that use Kalman-filters to directly estimate the object state from image features, the proposed method has a variety of advantages: Computation time is only O(n) rather than O(n/sup 3/) where n is the number of image features considered, sensor fusion is simplified and temporal estimation is decoupled from the choice of image features. The last point is of particular importance if occlusions that may occur during tracking are to be predicted and dynamically handled. With the tracking method proposed, a robot could be precisely controlled with respect to static objects and robustly follow targets moving in 6 degrees of freedom, while occasions were continuously predicted and appropriate features automatically selected at video rate (25 Hz). High robustness is obtained by Hough transform-based feature extraction.

This paper deals with the problem of obtaining a rough estimate of three dimensional object position and orientation from a single two dimensional camera image. Such an estimate is required by most 3-D to 2-D registration and tracking methods that can efficiently refine an initial value by numerical optimization to precisely recover 3-D pose. However the analytic computation of an initial pose guess requires the solution of an extremely complex correspondence problem that is due to the large number of topologically distinct aspects that arise when a three dimensional opaque object is imaged by a camera. Hence general analytic methods fail to achieve real-time performance and most tracking and registration systems are initialized interactively or by ad hoc heuristics. To overcome these limitations we present a novel method for approximate object pose estimation that is based on a neural net and that can easily be implemented in real-time. A modification of Kohonen's self-organizing feature map is systematically trained with computer generated object views such that it responds to a preprocessed image with one or more sets of object orientation parameters. The key idea proposed here is to choose network topology in accordance with the representation of 3-D orientation. Experimental results from both simulated and real images demonstrate that a pose estimate within the accuracy requirements can be found in more than 81% of all cases. The current implementation operates at 10 Hz on real world images.

A new algorithm for registering a 3D model to a single or a sequence of 2D images without a-priori knowledge of correspondences is presented. The algorithm is composed of two iterated steps: the first step determines correspondences between image features and model data, while the second one computes the 3D rigid body transformation that minimizes the displacement of the matched points. As camera images do not yield full 3D information, the key idea proposed is to exploit the inverse perspective, i.e. the projection ray determined by the focal point of the camera and an image feature point. The inverse perspective yields the constraints for both correspondence search and pose estimation in 3D space rather than in the image plane. This not only simplifies the motion parameter estimation but also reduces matching ambiguities due to occlusions. In addition, a robust M-estimation technique reduces the impact of false correspondences. Experimental results show that complex, 3D CAD-models can be registered efficiently and accurately to images, even if image features are incomplete, fragmented and noisy.