3D Closed Loop Boundary Detection and 6 DOF Pose Estimation

For vision guided robotic assembly, one of the fundamental enablers is the robust estimation of 6 degree-offreedom (DOF) pose of industrial parts or subassemblies. In this paper, we present a method to estimate 6 DOF pose of automotive sheet metal panels using 3D closed loop boundary (CLB) features from a stereo vision. The 3D CLBs extracted are used to identify the corresponding CAD model and estimate its 6 DOF pose with reference to the camera frame. The novelty of the proposed method lies in the fact that 3D CLBs are extracted efficiently by matching 2D CLBs from the stereo pair with its search space confined to the region of interest (ROI) and by reconstructing only the 3D data of the matched CLBs using the epipolar constraint. Our proposed method of the 6 DOF pose estimation using 3D CLBs has been demonstrated and applied to several decklid inner panels at GM Research Lab. Experimental results indicate that the proposed method offer computation efficiency less than one second and high performance under occlusion: over success rate 90% under 15% of occlusion.