Development and Evaluation of Filtering Methods for Flash LiDAR Data with Probabilities

Flash LiDAR is one of the most prominent solid-state perception sensors in autonomous driving, robotics, and activity monitoring applications. Its power constraints result in low power laser. This low power laser leads to erroneous distance measurements due to the background light in the scene. The latest literature on Flash LiDAR data processing by Fraunhofer IMS has proposed a novel method that generates unprecedented probability information for the points in the LiDAR point cloud to improve range detection. However, in addition to the probability information, it generates many false points in the point cloud. We extend this research at Fraunhofer IMS through this thesis. We propose a multistage noise reduction method that leverages the new probability information and spatial correlations to improve the quality of the point cloud. Our method removes at least 99% false points.
The proposed method outperforms the conventional LiDAR data processing methods on the datasets used in this work.