LiDAR-based Road State Estimation

Abstract

Abstract This master thesis explores the potential of LiDAR sensors and machine learning techniques for classifying the road friction coefficient and surface condition of road surfaces. The primary focus is to investigate the significance of various features derived from LiDAR data, including the reflectance value. To enable the research, a dataset representing road state was collected, requiring the development of methods for extracting road points and accumulating point clouds over time. Furthermore, a multilayer perceptron and a transformer network were designed and evaluated on the created dataset. The results indicate promising performance of both the multilayer perceptron and the transformer models in distinguishing between road surfaces with high and low friction coefficients, as well as accurately identifying surface conditions associated with each respective class. In this study, dry and moist road surfaces are together associated with the high friction coefficient class. Similarly, snowy and icy surfaces are grouped together and associated the low friction coefficient class. The achieved balanced accuracy and F1-score reached 77.7% and 74.6% when discriminating between high or low friction coefficients, as well as 77.1% and 74.6% when estimating surface conditions within each friction coefficient class. However, when considering the surface condition classes individually, the models could not make accurate classifications. Moreover, the best result was obtained using the reflectance value of LiDAR points in combination with the distances to them. The findings of this research contribute to the understanding of the reflectance value of LiDAR points on various surface conditions. Additionally, the study revealed that utilizing single LiDAR points for estimations yielded adequate results, although a trend was seen where leveraging multiple points indicated an improved performance. The insights gained from this research can guide further advancements related to prediction of the friction coefficient and surface condition.