AI in public transport: advanced sensor-driven forecasting system in Vilnius and Seoul
We talked to our researcher Henrikas about a recent project, undertaken by AAI Labs in collaboration with DFRC Group, on advanced sensor-driven forecasting system for public transportation.
Could you please start by introducing the project?
This project tackled the challenge of forecasting passenger flow in Vilnius's public transport system, specifically buses and bus stops. In collaboration with the Korean-based partner DFRC Group, we focused on integrating advanced sensors into the transport infrastructure. The objective was to enhance scheduling efficiency and responsiveness to real-time passenger numbers at the bus stop. After an extensive data collection phase, we crafted several machine learning models to predict passenger flows accurately.
Can you tell us more about the problem that required a new solution?
The issue of passenger flow prediction in public transport isn't new, but our approach is. By utilizing novel sensor technology, we aimed to surpass the accuracy of existing methods: WiFi-based mobile phone detectors and bus door scanner systems.
Can you elaborate on your personal view of how it affects companies?
From my perspective, the passenger flow prediction models we developed hold significant value for public transport providers. These models offer a new level of insight into passenger behavior, enabling more efficient scheduling and resource allocation. Not only does it optimize operational costs but also enhances the passenger experience by reducing wait times and overcrowding in public transport stations.
Do you see the potential for this product to be used outside of test sites? Can it be adapted elsewhere?
The success of this project in Vilnius and Seoul opens up exciting possibilities for other regions. The hardware we used is versatile and can be implemented in various locations, provided that sensors have adequate power supply and environmental protection. The adaptability of our system lies in its ability to be calibrated to different urban transit environments, allowing for broad application potential beyond just our initial test cities.
What was the greatest achievement of this project?
A personal highlight was engaging hands-on with the hardware used for data collection. Transforming this real-world data into effective machine-learning models, and then visualizing these insights, added substantial depth to the project. Collectively, we're proud of creating a comprehensive system in a relatively short time frame.
Were there any risks associated with the project, and how did you mitigate them?
Like many machine learning initiatives, the primary risk was data adequacy. The challenge of installing sensors and navigating bureaucratic hurdles raised concerns about collecting sufficient data on time. We tackled this by initiating model development with limited real data, supplemented by generated data. This approach proved successful, allowing us to meet our accuracy targets.
For more information about the project, see the publication here.
