Closing the gap in sidewalk data
Sidewalks (a.k.a footpaths) have been a much-overlooked piece of infrastructure in many cities until very recently. The growing number of people walking, cycling, and using emerging micro-mobility modes (e.g. e-scooters) in a world hit by a pandemic has made governments realize the critical role that the active transportation infrastructure plays in urban liveability and resilience.
The emergence of automated delivery robots (e.g. Starship Technologies, CoCo, Serve Robotics, Kiwibot) and their primary use of footpaths have put the walking infrastructure in the spotlight even more. It has also created a previously missing commercial incentive for better mapping of sidewalks.
Many cities don’t even have complete and regularly maintained routable walking network maps that can be used for navigation purposes for humans and robots.
Data on walking infrastructure are quite scarce. Very few cities know where exactly their bike racks, garbage bins, or accessibility ramps are. Many don’t even have a good understanding of their footpaths' width and their connectivity and accessibility.
Sidewalks as complex multi-layer environments
Sidewalk networks are complex environments. Unlike roads with clear traffic lanes and directions, movements on sidewalks largely rely on pedestrians' unique self-organization capabilities. Pedestrians move in multi-directions. A wide range of static and dynamic objects can be found on sidewalks in various shapes and forms, from garbage bins, portable and fixed toilets, bike racks, benches, bus shelters, utility boxes, traffic control boxes, ramps, stairs, dogs, cats, birds, motorbikes, wheelchairs, scooters, and many more.
Graph layer
The most basic footpath data layer is perhaps the walking network layout. An article published by the Mapillary team in 2021 discusses the extent to which walking route data in OpenStreetMap varies significantly between cities.
The gap in sidewalk data can be revealed with a quick look at the OSM maps of Sydney and Chicago, the two cities that I consider home. The sidewalk network data on OSM for Chicago seems quite comprehensive and complete while there is barely any walkway information available forSydney.
Since 2020, Google Maps has been adding more street-level details such as crosswalks and sidewalks to their maps but for only a limited number of cities. Mapping all sidewalks globally is challenging and may take huge efforts, resources, and many years to complete.
At footpath.ai, we automatically generate a base sidewalk network layout using information such as road centreline, building footprints, and satellite images. We are planning to improve our mapping process using footpath-level imagery and semantic information to increase the accuracy and to estimate sidewalk attributes such as widths and edges.
Visual layer
Other than the network layout information, the visual information on footpaths is also very limited and scarce. Street-level images provided by companies like Google Maps or Mapillary are focused on roads and provide a “car-centric” view of streets.
At footpath.ai, we aim to fill this gap by building a global 360-degree footpath-view imagery dataset that is snapped to the sidewalk network layout, much like Google Street View but for sidewalks only. This provides a “pedestrian-centric” or “delivery robot-centric” view of streets, depending on the application.
Semantic layer
We are also taking a big further step, extracting semantic information from our collected 360-degree footpath-view images using deep learning.
The first task is to build a large-scale manually annotated and fine-grained semantic segmentation training dataset. Several publicly available training datasets for semantic segmentation of street scenes already exist such as Cityscapes, Mapillary Vistas, COCO, and ADE20K. However, they are focused on roads with a “car-centric” view. Whether these datasets enable accurate scene understanding on sidewalks is a big question and the motivation behind what we are building at footpath.ai. Our 360-degree footpath-view semantic segmentation training dataset will enable unique urban feature extraction with a focus on sidewalks for pedestrian navigation and autonomous delivery robots' scene perception.
What’s next?
We still have a long way to go. We’re aiming to launch our sidewalk data map platform and APIs over the next 6–12 months. We’re hoping to collect data from as many cities as possible across the world, from cities down under in Australia and Asia to Europe, Africa, and America.
