BOOSTING U-NET WITH FOCAL LOSS FOR ROAD MARKING CLASSIFICATION ON SPARSE MOBILE LIDAR POINT CLOUD DERIVED IMAGES

Abstract. Road markings play an important role in vehicular navigation. It helps provide sufficient information for safe driving and smooth traffic flow. As such, with the rise of digital maps such as High-Definition (HD) maps, which are used by autonomous vehicles or self-driving cars, they must be well represented in their digital counterparts. However, survey-grade mobile mapping systems are expensive and thus open the idea of using lower-cost/level LIDAR sensors for mapping. Unfortunately, using such sensors provide sparser point clouds. This work aims to propose a method that successfully classifies road markings on sparse mobile LIDAR point cloud-derived images using UNET trained with focal loss. Results have shown successful road marking classification with a 94.68%increase in recall and a maximum 49.39% increase in F1-score. Adjusting precision by removing the insignificant class (“black”) further increases the resulting F1-score to 82.74%. Extending the method produces a classified point cloud by combining the classified image with a depth image. This research also aims to help aid boost the research on lower-cost/level sensors for mobile mapping purposes.