Detection and Pose Estimation of Top Circular Objects Using Mask R-CNN and Overlap Handling Algorithm

This paper proposes an algorithm for detecting and estimating the pose of top objects in a complex environment where thin metal circular plates are randomly stacked. In complex environments where multiple instances of the same object are randomly stacked, the robot needs to detect and compare objects to identify the top ones for grasping. Our approach involves a combination of deep learning-based instance segmentation and an overlap handling algorithm for precise top object detection. Subsequently, leveraging three-dimensional geometric data, we estimate the object's pose by determining its plane. To validate the proposed algorithm, we constructed two environments consisting of objects with different sizes and thicknesses. The first experiment quantitatively validated the object detection and overlap handling algorithm. The second experiment quantitatively compared different plane estimation algorithms. The third experiment quantitatively compared the pose of objects using the G-ICP (Generalized Iterative Closest Point) algorithm and the proposed algorithm against the ground truth pose. Additionally, we performed a qualitative comparison by visualizing the poses estimated by each algorithm in the images. In the experimental results, the overlap handling algorithm had an average success rate of 84.21%. Additionally, pose estimation using G-ICP before plane estimation frequently resulted in issues like drift in the center point and frequent misalignment with areas other than the object. On the other hand, pose estimation using G-ICP after plane estimation and the proposed algorithm yielded similar performance with average ADD-S values of 6mm or less. However, the pose estimated using the proposed algorithm resulted in a minimum 0.25x reduction in execution time compared to the G-ICP algorithm.