3D Pose Estimation for the Object with Knowing Color Symbol by Using Correspondence Grouping Algorithm

Abstract

Listen

Translate article

Color feature matching for point cloud data with color 2D code is proposed to recognize the object whose unique feature is difficult to be made by only using shape information. When a specific object such as a commercial product is recognized, it is possible to inquire the product information by attaching a known marker registered in the network server. However, even when a known marker is attached, it is difficult to acquire the accurate marker position and orientation unless the image is taken from a direction that allows good recognition. On the other hand, the position / orientation estimation method using point cloud information cannot be effective unless the shape information is easy to appear. In the proposed method, in order to obtain product information and accurately measure the position and orientation of the product, a marker that is given a characteristic color is attached to a known location on the target object so that the color information as well as shape information can be evaluated simultaneously. In the case of packaged products, it is not necessary to give color information in particular. However, the artificial known marker could be the important evaluation item to estimate object posture since the products that are covered with a transparent film such as a sandwich and a rice ball in the convenience store have different textures for each individual. In order to evaluate the accuracy of the proposed method, the recognition results are compared in a 3D virtual environment. For the evaluation of matching with corresponding grouping algorithm, two major feature-based matching algorithm are implemented; one is 3D-HV (Hough Voting) which returns acceptable inlier searching performance in real-time, and another is RANSAC (Random Sample Consensus) which is known as the high precision performance. In addition, 2.5D and complete 3D reference model are prepared to compare the computational time and the accuracy of the posture estimation. As a result of the experiment, the method using the CSHOT feature, the RANSAC and the complete 3D model as the search method has the highest accuracy and is practical as the search time.