Supervised pose and motion estimation over weakly textured industrial objects

Visually estimating the motion of automotive body parts over an assembly line represents a major challenge for classical feature detection, matching and tracking algorithms due to the lack of a rich surface texture. But as feature extraction and matching remain vital for accurate object pose and motion estimation, this paper presents a thorough investigation on the actual reliability of popular feature extraction and matching tools in terms of stability and robustness for industrial applications. Severe tracking errors that result from brightness variations and occlusions are corrected with the integration of an original supervisory approach that relies on the encoding of a minimum amount of a priori information about the general appearance of the objects. The proposed solution is experimentally validated on an application for quality control in the automotive industry.