Design of dust-filtering algorithms for LiDAR sensors in off-road vehicles using the AI and non-AI methods

Abstract

The performance of Lidar sensors degrades in the presence of dust. These particles can impact sensor measurements and cause robot perception algorithms to misinterpret data. This thesis proposes two distinct dust filtering methods to address this issue. These methods utilize both AI and non-AI techniques. Specifically, we designed various dust filters including the Low-Intensity Dynamic Outlier Removal (LIDROR) using intensity and range information. In addition, we proposed a voxel-based classification method with multiple classifiers, such as Random Forest (RF), Support Vector Machine (SVM), and Deep Neural Network (DNN). Two dust LiDAR datasets were collected and labeled for evaluation purposes. All proposed algorithms were implemented in the Robotic Operating System, allowing for the testing of these filters in real time. Using labeled data, a comprehensive comparison was made between these two methods. The proposed filters outperform conventional filters in terms of achieving dust removal without losing the surrounding data.