QR-Based Standing Risk Guidance System for Passengers with Mobility Needs
Predicting standing likelihood and time to seating for more accessible bus travel.
Overview
Developed as part of the TS × KT Digital Talent Scholarship social impact project, this service helps passengers with mobility needs make informed bus travel decisions.
Users scan a QR code at a bus stop and enter their destination to compare available routes by travel time, standing likelihood, and estimated time to seating.
The system combines LightGBM prediction models with real-time bus information to estimate the standing burden of each route and provide guidance through a web interface.
Problem
Standing on a moving bus can place a substantial physical burden on older adults and other passengers with mobility needs.
Current crowding information helps passengers assess conditions before boarding, but it does not indicate whether they will find a seat or how long they may need to stand.
The project addresses this gap by predicting standing likelihood and time to seating, allowing passengers to consider seating availability alongside travel time when choosing a route.
My role
- Developed LightGBM classification and regression models to predict standing likelihood and time to seating.
- Evaluated the models across 11 monthly test windows, achieving a mean classification AUC of 0.9175 and a mean regression MAE of 157.2 seconds.
- Applied temporal subsampling to reduce training time by 55% while retaining 99.8% of the AUC achieved with the full training dataset.
Approach
- The prediction pipeline uses 2024 transit card records from Seongbuk-gu, Seoul, to reconstruct passenger occupancy along bus routes.
- Boarding and alighting records are combined with estimated seat capacity, and a first-in, first-out seating simulation generates labels for standing status and time to seating.
- A LightGBM classifier estimates whether a passenger will need to stand after boarding, while a separate regressor estimates the time until a seat becomes available.
- These predictions are combined with real-time bus information to compare routes serving the user's destination.
System architecture
- Access: A QR code identifies the boarding stop and opens the service, where the user enters a destination.
- Route retrieval: Bus location and arrival APIs provide information on available routes and approaching vehicles.
- Prediction: LightGBM models estimate standing likelihood and time to seating for candidate journeys.
- Guidance: The interface presents travel time, predicted standing burden, and crowding information to support route selection.
- Application: A React frontend, Spring Boot backend, and MySQL database support the service.