D-Fast-SCNN + Combo Loss: Improved Road Marking Extraction on Mobile Lidar Sparse Point Cloud-Derived Images

Recent works have explored the automatic extraction of road marking features from sparse point clouds using convolutional neural networks (CNN) in an attempt to support the use of low-cost light detection and ranging (LiDAR) sensors for reducing the cost of mobile mapping tasks. In this work we follow through with this ideology and propose D-Fast-SCNN, a modification of the Fast-SCNN model, to improve its classification accuracy while still leveraging its speed. We propose the addition of a dilation block, which utilizes the morphological process of dilation. This additional process takes full advantage of the observed property of sparse point cloud-derived imagery, wherein misclassifications can be omitted in the areas with no corresponding point cloud value. After multiple trials, we can see an increase of around 10% in the mean f1-score and a decrease of roughly 6% in uncertainty for the target road marking class. In future works an ablation study will be conducted to investigate its effects when employed elsewhere in the model.