Project

DESCRIPTION:

Point clouds are discrete data points representing 3D shapes and objects in space. Thanks to its ability to accurately describe complex spatial information, point cloud emerged as an attractive technique for various applications, including building modeling, infrastructure mapping, autonomous driving, robotics, virtual reality, etc. However, due to imperfections of measurement instruments (e.g. Lidar or laser scanner) and varying conditions of the measurement environment, raw point cloud data tend to be noisy, which negatively affects the accuracy of the subsequent tasks. Noise in point cloud data captured from large complex scenes, e.g. point cloud of a street captured by a mobile mapping system, is more complex and harder to remove, since various factors, such as material properties (e.g., color, sheen, etc.), object distance, and laser incidence angle, impact noise characteristics. This project investigates a two-step approach to address this challenge. It first segments the 3D space based on semantics or distance and then adaptively applies customized denoise engines to each segmented region.