Improved Point-Cloud Global Registration with Segmentation and Ray-Casting Model

Abstract

This thesis proposes a novel point cloud registration method based on Segmentation and Global Registration. Because the point cloud captured by the depth camera includes not only the point cloud of the target object but also the point cloud of the background, registration between the model point cloud and the captured point cloud often leads to undesired results. Another problem is that the excessive number of points from the background brings a lot of computation, which greatly affects the point cloud registration time and point cloud registration efficiency, which makes the point cloud registration difficult to complete. To solve this problem, I propose using segmentation. By applying the segmentation network on the point cloud, we can remove a lot of noise data from the background point cloud. When the point cloud captured by the depth camera, not by the 3D scanner is being used in registration, there is a problem. The two models are different. The 3D scanner can capture almost all the points on the whole surface of the model, but the depth camera can only capture the point on the side facing the depth camera. The difference is quite big affecting the registration performance. Therefore, the CAD model is resampled by the Ray-Casting method to create a point cloud similar to the point cloud captured by a depth camera. For the 5 objects, 900 times registration was done without the segmentation, and another 900 times registration with the purposed improved method, in a total of 1800 times. When comparing the two methods, the proposed method increased the fitness by 4.31%, the recall by 4.01%, and reduced the root mean square error (RMSE), distance error, and angle of the normal vector by 3.65%, 18.50%, and 22.83%, respectively. The time for one registration was also reduced by 58.25%. In addition, machine learning based judgment is proposed. In the experiment, we observed that the single measure such as fitness can easily fail to judge the success or the failure of the registration. Some combination of multiple measures also could not solve this problem. Therefore, a deep learning network using multiple measures as inputs are designed and learned used to evaluate the registration result. The final accuracy reached 91.67%(1650/1800). The proposed method not only improves the speed of global registration but also enhances the success rate of global registration, which provides the possibility of 6D pose estimation.