Registration of 3D Point Clouds with Credibility for Changing Environments.

In progress of technologies related to augmented reality and autonomous robot mobility, needs of registration techniques that use a pre-made global 3D point cloud as a map and presently-acquired local one for localization increases rapidly. However, in a warehouse of shops and restaurants, location of objects such as products and food materials changes very frequently, resulting in more difference between the pre-made environmental map and the local data of 3D point cloud. This difference reduces accuracy of localization. One of problems in the localization is that when is acquired in both point clouds, the object is used as a reference, resulting in an error in the overall registration accuracy. Therefore, in this paper, the environment point cloud is updated at each registration to reduce the difference from the local point cloud. The updated information is also used to find reference points during registration. Specifically, point cloud data with general coordinate values is made into point cloud data with attributes by recording the past acquisition history. From that acquisition history, a value called Credibility is calculated, which evaluates the likelihood that the point can be used as a reference point during estimation. Then, based on Credibility, I propose a positioning method that updates the environmental point cloud and reduces the impact from moving objects. In the experiment, after updating the environmental point cloud using point clouds acquired at different times by changing the placement of objects on the desk and adding the acquisition history, the positioning accuracy with the point cloud acquired by moving the objects again was evaluated using the squared error. The evaluation results showed approximately 30 percent increase in positioning accuracy to the proper location based on a non-moving object compared to the existing ICP algorithm.