3D Pose Estimation of Bin Picking Object using Deep Learning and 3D Matching

In this paper, we propose a method to estimate 3D pose information of an object in a randomly piled-up environment by using image data obtained from an RGB-D camera. The proposed method consists of two modules: object detection by deep learning, and pose estimation by Iterative Closest Point (ICP) algorithm. In the first module, we propose an image encoding method to generate three channel images by integrating depth and infrared images captured by the camera. We use these encoded images as both the input data and training data set in a deep learning-based object detection step. Also, we propose a depth-based filtering method to improve the precision of object detection and to reduce the number of false positives by pre-processing input data. ICP-based 3D pose estimation is done in the second module, where we applied a plane-fitting method to increase the accuracy of the estimated pose.