Point Cloud Registration with Preprocessing Enhanced Robustness for Structured Light Sensors

Structured light sensors can obtain dense point clouds with high precision at close range, but also introduce unfavorable factors including random fragments consisting of several points, uneven distribution of points, as well as multiple boundaries. These factors lead to erroneous feature extraction and matching when using the existing feature-matching registration methods, especially methods based on M-like estimators which have an efficiency advantage and do not rely heavily on initial estimations. Therefore, we propose an improved registration method together with preprocessing to reduce the impact of these factors, thus enhancing the robustness. According to experiment results on our custom dataset from a real-life scene, the proposed method shows a minimum decrease of 21.1% in translation root mean square error (RMSE) and 35.7% in rotation RMSE compared to existing methods without losing efficiency.