Mapping Pedestrian Ways with Computer Vision and Street-level Imagery
2021-10-01, 10:30–11:00, Ushuaia

Mapping pedestrian infrastructure on OpenStreetMap has entered a new age. With the latest AI-powered mapping tools from Facebook, sidewalks and crosswalks visible in Mapillary imagery can now be easily translated into geospatial data, representing the network of routable features on the map.

Sidewalks are historically under mapped in OpenStreetMap, with many towns and cities worldwide showing only roads. For the pedestrian, adding sidewalks to the map, as well as using a map with accurately displayed footways, is a frequent challenge. This is all changing with a new workflow that uses computer vision to derive sidewalks and crosswalks from Mapillary street-level imagery and predict their geospatial locations.

Join us in this session to learn more about how we created a sidewalk and crosswalk network from street-level imagery, and how to use the latest tools to map a sidewalk near you.

At Facebook, we are working on the latest AI-powered mapping tools to help generate a large scale open dataset of sidewalks, derived entirely from user contributed street-level images to Mapillary. This project focuses on three main aspects:

1) an overarching goal of turning user contributed imagery into user-contributed sidewalks and crosswalks
2) solving the challenge of how exactly a street-level imagery and computer vision detected sidewalks can be projected onto the map as sidewalk and crosswalk data
3) presenting the OpenStreetMap community with a robust, practical, and well-crafted tool for creating maps with this sidewalk and crosswalk data

The role of the community is key in both being able to capture Mapillary imagery as well as map sidewalks derived from their own imagery. The community can provide both the input and the output of this workflow. The more imagery a community captures, the more the available sidewalk data scales in previously unmapped areas.

The Facebook team has focused on finding the best method of converting street-level imagery from Mapillary contributors to a spatial network of line data. Many steps are involved in order to provide clean data that accurately indicates where sidewalks and crosswalks may exist, and where human validation can be applied. In addition, the Facebook team has developed the most effective interface and workflow for working with this data, truly empowering a mapper to add new map data with quality and scale like never before.

Authors and Affiliations

Beddow, Christopher (1)

(1) Facebook Reality Labs, Mapillary


Use cases & applications


New trends: IoT, Indoor mapping, drones - UAV (unmanned aerial vehicle), Artificial intelligence - machine learning, deep learning-, geospatial data structures, real time raster analysis


1 - Principiants. No required specific knowledge is needed.

Language of the Presentation


Christopher Beddow is an analyst focusing on Mapillary and OpenStreetMap on the Spatial Computing team at Facebook Reality Labs. He has worked on the OSM, GIS, and map data products at Mapillary since 2016. He codes in Python, SQL, and JavaScript, and enjoys traveling widely while editing OpenStreetMap.