Development of a multi-stage 3D object recognition system for bin-picking based on a geometric framework using a low-cost RGB-D camera

Fast and accurate recognition of the position and pose of mechanical parts in 3D space is an important step in bin-picking tasks. In this research, we have developed an effective recognition system that can detect many parts from a stacked state in one detection attempt with a low-cost RGB-D camera. The system is designed with a multi-stage recognition pipeline. At first, an AI-based coarse identification that quickly determines the area of prioritized pose parts in the supply box (bin), which is used to create a region of interest (ROI) that narrows down the scope for detailed identification in the next stage. Next, a new geometric feature-based parts recognition method is applied, which is named Fit-by-Appearance. This core method can perform fast and accurate multi-step recognition within the ROI. Finally, the recognition results from the fit-by-appearance stage are used to further improve the accuracy of final results by applying the point cloud ICP algorithm. The proposed recognition method has been validated using industrial parts. It is capable of identifying a large number of parts from randomly stacked state in a single attempt, even if these parts are occluded from each other. The detection time of approximately 100 ms.