Noise Removal from Point Clouds through Sensor Fusion of LiDAR and Camera

Abstract

In this study, the method of noise removal in 3D point clouds using sensor fusion of LiDAR and camera was introduced. First, a point cloud measurement was performed using Matterport Pro3 in Fukushima Prefecture, Japan. The bridges and other man-made objects were scanned while people were also scanned. The relative positions of the camera and LiDAR were adjusted using a checkerboard for cross-calibration. Internal and external parameters were then obtained to map the 2D image to the 3D point cloud. This allowed us to correlate each point in the point cloud with a pixel in the image and develop a segmentation method for noise removal. Next, we extracted the region of a person in a 2D image and mapped it to a 3D space to classify noise points related to the person in the point cloud. It was confirmed that this method can effectively classify people even in the presence of many other objects in the point clouds. The evaluation results using recall, precision, and F1 scores had mean values of 0.923, 0.878, and 0.889 for all samples, respectively, indicating that highly accurate noise reduction is possible. The results of this study are expected to be effective for cleaning and pre-processing 3D point clouds. Future perspectives include evaluating the accuracy of this method on data with a larger number of people.