Pose Estimation for Variable Configuration Objects: an Evolutionary Approach to Vision-based Navigation and Inspection

摘要

This paper concerns vision-based relative navigation of autonomous vehicles for satellite servicing. The task is twofold: first, the target object's relative position and orientation (pose) is to be determined, and second, a visual inspection is to be performed. An evolutionary-based algorithm for the problem of pose estimation of 3D objects using 2D camera images has been developed. The procedure consists on looking for six position parameters, three for rotation and three for translation, such that the projection of a model best fits a set of points (vertices) extracted from a 2D image. A population of candidate solutions is used, and their goodness is measured in terms of distance between the model and image points. A distinguishing feature of the proposed algorithm is that the best matching model is also looked for. This feature is used for visual inspection in order to detect macro defects. In addition, important features of the algorithm include speed and robustness with respect to noise on the input data. Experimental results conducted both on synthetic and real images demonstrate the effectiveness of the proposed approach.