A fast and accurate point cloud registration network for industrial point cloud data

Abstract

The point cloud registration task for anomaly detection in industry has the requirements of minimal inference time and maximal performance by using the industrial point cloud data which has the characteristics of high precision, large number of points in a single point cloud, and easy to be affected by noise. In this work, we curated a laptop dataset for testing our network on the real industry situation. We propose a sampling-registration combination, SampleNet-iPCRNet, to fast and accurate to perform the point cloud registration task in industry. The SampleNet and an iterative PCRNet are cooperated to simultaneously sample a set of representative data, to enhance the noise resistance, and to perform the point cloud registration task. As a result, the SampleNet-iPCRNet achieves 0.9849 in AUC, 4.1054 in rotation error, 0.0021 in translation error, and 182 ms in inference speed which outperforming among other sampling-registration combination strategies. Results show that the proposed method can effectively improve the registration performance and speed on the dataset representing the real situation in the industry. Supplementary information can be available at: https://pan.baidu.com/s/1CwqSb_oMRrdN5x3ipAYfuQ with the jb0y as the extraction code.