6D Pose Estimation for Texture-Less Industrial Parts in the Crowd

Recovering the 6D object pose of industrial part is a common but challenging problem in many robotic applications. In this paper, an accurate 6D pose estimation approach is proposed for texture-less industrial part in the crowd. The proposed method consists of three stages: object detection, pose hypotheses generation, and pose refinement. Firstly, the bounding boxes of object instances in an RGB image are detected by a convolution neural network. The training dataset is automatically synthesized using an efficient image rendering method. Then, highlight detection and removal are employed to eliminate noise edges. The coarse pose hypotheses are generated using an edge-based fast directional chamfer matching algorithm. After that, the accurate 6D poses are obtained by applying a non-linear optimization to these pose hypotheses. A re-weighted least-squares loss function is utilized to suppress outlier noise in optimization. Finally, an edge direction consistency score is used to evaluate these obtained poses and eliminate outliers. The proposed method only relies on single RGB image to recover the 6D object pose in the crowd. Experimental results of texture-less industrial parts show the accuracy and robustness of the proposed method.