User assisted and automatic inverse procedural road modelling at the city scale

TL;DR

This paper presents a fast, robust method for automatically refining coarse city-scale road models using various observational data and user input, significantly improving accuracy for urban mapping.

Contribution

It introduces a simple, fast optimization approach that enhances initial road models with diverse data sources, achieving high accuracy and enabling practical city-scale applications.

Findings

Median distance reduced from 1.5m to 0.45m automatically

Optimization runs in a few minutes for entire city, seconds for small areas

Method is robust, fast, and adaptable to various data sources

Abstract

Cities are structured by roads. Having up to date and detailed maps of these is thus an important challenge for urban planing, civil engineering and transportation. Those maps are traditionally created manually, which represents a massive amount of work, and may discard recent or temporary changes. For these reasons, automated map building has been a long time goal, either for road network reconstruction or for local road surface reconstruction from low level observations. In this work, we navigate between these two goals. Starting from an approximate road axis (+ width) network as a simple road modelling, we propose to use observations of street features and optimisation to improve the coarse model. Observations are generic, and as such, can be derived from various data, such as aerial images, street images and street Lidar, other GIS data, and complementary user input. Starting from an initial road modelling which is at a median distance of 1.5 metre from the sidewalk ground-truth, our method has the potential to robustly optimise the road modelling so the median distance reaches 0.45 metre fully automatically, with better results possible using user inputs and/or more precise observations. The robust non linear least square optimisation used is extremely fast, with computing time from few minutes (whole Paris) to less than a second for a few blocks. The proposed approach is simple, very fast and produces a credible road model. These promising results open the way to various applications, such as integration in an interactive framework, or integration in a more powerful optimisation method, which would be able to further segment road network and use more complex road model.