Finding crosswalks, tactile pavings and sidewalk widths with AI ~ Suresh Devalapalli & Ricky Zhang

Summary

In this presentation, Suresh Deopali and Ricky from Gshan Solutions, in collaboration with the University of Washington, discuss their AI-driven work mapping pedestrian infrastructure across the state of Washington. They focus on three key extensions: classifying crosswalk types, detecting tactile paving, and estimating sidewalk widths. Using AI, they have automated the process of mapping over 36,000 kilometers of sidewalks, 6,700 kilometers of crossings, and 133,000 intersections. The project aims to improve accessibility and mapping efficiency, addressing challenges such as image resolution and occlusions. They encourage local communities to collaborate and contribute to the OpenStreetMap (OSM) project.

Takeaways

• 😀 AI is being used to map pedestrian infrastructure across Washington state, which covers 184,000 square kilometers.
• 😀 The project involves classifying crosswalk types, detecting tactile paving, and estimating sidewalk widths using AI.
• 😀 The team developed an AI extension to classify crosswalks as marked or unmarked and further categorize marked crossings into zebra or other types.
• 😀 8,000 images were used to train the crosswalk classification model, achieving an F1 score of 0.84 on unseen data.
• 😀 Challenges in crosswalk classification include shadow and tree cover, which hinder the AI's ability to classify crossings accurately.
• 😀 Tactile paving detection was carried out using 4,000 image patches around curb nodes, employing ensemble learning techniques for higher accuracy.
• 😀 The tactile paving detection model has an 80% accuracy rate but struggles with low image resolution and shadow coverage.
• 😀 Sidewalk width estimation was automated using aerial imagery and semantic segmentation, with an error margin of less than 0.5 meters (3 pixels).
• 😀 The sidewalk width estimation helps improve accessibility for users, including wheelchair users, by identifying narrow sidewalks.
• 😀 The project seeks to collaborate with local communities to expand data for OpenStreetMap (OSM) and improve the AI's data set.
• 😀 Future improvements include addressing occlusion challenges, expanding crosswalk classification categories, and incorporating more diverse data sources.

Q & A

What is the primary goal of the AI system developed by Gshan Solutions and the University of Washington?

The primary goal is to automate the mapping of pedestrian infrastructure across the state of Washington, including crosswalk types, tactile pavings, and sidewalk widths, using AI technology.

How does the AI classify crosswalk types?

The AI classifies crosswalks by first identifying if they are marked or unmarked using aerial imagery. For marked crosswalks, it further divides them into zebra crossings and other types, using a classification model trained on 8,000 images.

What challenges does the AI face in accurately classifying crosswalk types?

The AI faces challenges with occlusions, such as tree cover and shadows, which can obstruct visibility of the crossing and affect classification accuracy. These issues are being addressed by incorporating additional data sources like street imagery.

How is tactile paving detected by the AI?

Tactile paving is detected using a binary classification model trained on 4,000 image patches sampled around curb nodes. The model determines whether tactile paving is present or not at these nodes.

Why is tactile paving detection more challenging than crosswalk classification?

Tactile paving detection is more challenging due to the need for granular detail in the imagery. Low-resolution images and variations in the paving's appearance make it harder for a single model to confidently predict its presence, requiring ensemble learning to improve accuracy.

What is ensemble learning, and why is it used in tactile paving detection?

Ensemble learning is a method where multiple smaller, weaker classifiers are combined to make a final prediction. It is used in tactile paving detection to improve accuracy by taking the majority vote of different models, such as random forests and support vector machines, which helps overcome individual model limitations.

How is sidewalk width estimated using AI?

Sidewalk width is estimated by using semantic segmentation on aerial imagery to detect the sidewalk areas. Then, the width is calculated based on geographic coordinates and a kernel density estimation is used to determine the most common width along the sidewalk.

What is the mean error margin for sidewalk width estimation, and why is it important?

The mean error margin for sidewalk width estimation is less than half a meter. This accuracy is critical for accessibility applications, as users, especially those with disabilities, need to know whether sidewalks meet minimum width requirements for safe navigation.

What steps are being taken to address challenges with image quality, like low resolution and shadow coverage?

To address image quality issues, the team is incorporating additional data sources, such as street-level imagery, to improve classification accuracy. They are also expanding their dataset to include more diverse and out-of-distribution samples.

Is the data produced by this AI system publicly available?

Yes, the data is publicly available through OpenStreetMap (OSM) with a sign-up required for access. However, the data must be manually verified by local communities before being uploaded to OSM.