Lidar-based detection of airborne particles for robust robot perception

Abstract

Airborne particles such as dust, smoke and fog have a significant detrimental impact on Lidar- based robotic perception systems. Lidar rays can reflect on these particles, leading modern perception methods to erroneous results, such as false obstacles or misclassified elements. We propose a method to detect airborne particles in 3D Lidar point clouds using classification from geometric features and Lidar intensity re- turns. We compare three different classifiers and evaluate our approach on dust and fog data collected in outdoor scenarios. We achieve an accuracy of up to 95% in detecting airborne particles in Lidar point clouds, making our pro- posed method a promising solution for appli- cations such as obstacle detection and object recognition in outdoor environments. We make available the code and data for this work at https://leo-stan.github.io/particles detection.