A multi-feature fusion-based pose tracking method for industrial object with visual ambiguities

The robustness and accuracy of industrial object pose tracking is critical in manufacturing automation. Vision-based object tracking methods estimate poses by establishing feature correspondences between the object model and RGB images. However, due to the visual ambiguities caused by textureless surface, symmetric structure, color similarity to background, motion blur and occlusion in a complex environment, establishing feature correspondences for industrial object pose tracking remains challenging. In this paper, we propose an industrial object tracking method based on multi-feature adaptive fusion to improve the robustness and accuracy with visual ambiguity interference. Specifically, the degree of visual ambiguity of industrial objects is measured quantitatively to specify the complexity of industrial environments. Considering the complex background and occlusion of industrial objects, an improved color statistic model and a geometric edge model are proposed to reduce the effect of visual ambiguity during establishing feature correspondences. Afterwards, an adaptive weight assignment mechanism based on real-time observation status is proposed to maximize complementarity of multi-feature fusion. Finally, a hybrid optimization strategy is designed through considering non-local search and local optimization to reduce the possibility of trapping the local minima. Experiments conducted on semi-synthetic and real datasets demonstrate that our method achieves higher performance compared to existing methods and shows great robustness against visual ambiguities.