Page 736 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

race distribution if possible to determine that each race presented in the dataset is equally
represented.

Use Case Status: The use case is in progress and part of a research project�

2�2 Benefits of the use case

Sustainable Cities and Communities: Public transit lies at the heart of efficient urban mobility,
and keeping drivers alert is critical for accident prevention and rider confidence. By monitoring
and mitigating fatigue and stress in real time, our app makes buses, trams, and trains noticeably
safer—encouraging more commuters to choose mass transit over private cars and easing road
congestion. When passengers know operators are actively supervised for drowsiness, they feel
more secure, which is especially valuable for children, older adults, and people with disabilities
who depend on these services. In doing so, the solution not only reduces crash risk but also
fosters an inclusive network where everyone has equitable access. Finally, by boosting trust
and ridership, this technology helps cut carbon emissions and supports the long-term vision
of resilient, low-emission cities.

2�3 Future Work

Our proposed future work involves utilizing the app for data collection, to collect frames of the
driver alongside physiological data, such as heart rate obtained from wearables, while driving.
By gathering real-world data, we aim to enhance the accuracy and reliability of drowsiness
detection. Additionally, we plan to enhance this dataset by ingesting information from already
existing ones, to develop a more robust drowsiness detection model.

To ensure the proposed drowsiness detection system is both effective and equitable, we will
adopt a multi-dimensional evaluation framework that includes metrics for accuracy, fairness,
and latency. Accuracy will be measured using standard classification metrics such as precision,
recall, F1-score, and area under the Receiver Operating Characteristic( ROC) curve Area Under
the ROC Curve (AUC) on both general and user-specific datasets. Fairness will be evaluated
through the use of tools like Fairlearn, which provides detailed metrics such as demographic
parity, equal opportunity difference, and calibration curves across sensitive attributes (e.g.,
race, gender, facial features). These tools will be integrated directly into our model training
pipeline for ongoing bias monitoring and mitigation.

In addition latency, which is critical for real-time mobile application deployment, will be assessed
by measuring inference time on representative mobile hardware. The goal is to ensure sub-
second predictions (<500 ms) to allow timely intervention without requiring constant cloud
connectivity. Together, these validation methods will provide a robust and holistic view of
model performance in real-world conditions.

A key focus of our research is ensuring that the dataset is as diverse as possible. This diversity is
essential to eliminating racial and facial characteristic biases that often affect AI-driven detection
systems. We plan to train different models on a wide range of facial structures, skin tones, and
physical attributes, to create an unbiased system that provides fair and accurate results for all
users. This approach ensures that drivers of different ethnic backgrounds are equally protected,
promoting inclusivity in road safety technology.

700