An Effective Randomized Hough Transform Method to Extract Ground Plane from Kinect Point Cloud

Abstract

This article presents a method to improve plane segmentation on a 3D point cloud acquired and converted from Kinect data. This method utilizes Randomized Hough Transform (RHT) combined with angle and distance constraints to eliminate non-ground elements. The algorithm randomly selects 3 points to form a plane model, then estimates the surface normal on one of the three points to compare with the selected model and ensure the model fits a present plane in the point cloud. The planes that satisfy the predefined angle and distance constraints are selected out from all plane models found. Finally, all valid ground points which satisfy the best ground plane model can be found. The test results show that the elapsing time of the proposed modified RHT method is about to 10% of that running with the RANSAC algorithm and it is negligible in comparison with original RHT, while the benchmark completeness, correctness and quality are stable at high level. The result of this work is promising to apply for indoor robot navigation application using 3D computer vision.