A Feasibility Investigation of Bike Lane Width Estimation on LiDAR Point Cloud Using Dimensionality Reduction and Weakly Supervised Learning

Bike lane width is critical information for civil infrastructure planners to promote city sustainability, health, and urban livability. This research presents a feasibility investigation to develop a workflow for the recognition of bike lane markings using city-scale point cloud data. City-scale point cloud data was collected employing a backpack mobile scanner and an electric scooter in a US Southwestern City. To efficiently process and extract features automatically in this large city-scale data set, this study employed dimensionality reduction and weakly supervised learning to obtain information on bike lane widths. Dimensionality reduction involves projecting each road segment to 2D along the z-axis (up/down). The projected 2D images were paired with corresponding noisy annotations for bike lane markings and road curbs to train a weakly supervised learning model. The neural network outputs were postprocessed for bike lane width estimation. The output of this feasibility study shows that the produced workflow enables the automated recognition and localization of the bike lane markings. Ultimately, city-scale bike lane width characterization aims to support decision-makers in public infrastructure planning for increasing urban mobility and ensuring the safety and comfort of cyclists.