Page 737 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

Another crucial aspect of our development is optimizing the drowsiness detection model
for real-time performance, aiming to run it directly on the device, reducing reliance on cloud
processing. This will ensure faster predictions and immediate alerts, enhancing driver safety
without requiring constant internet connectivity. Real-time capabilities are critical in preventing 4.9: Accessibility
accidents, as immediate intervention can prompt drivers to take necessary precautions before
fatigue leads to dangerous situations.

By prioritizing fairness, efficiency, and real-time responsiveness, our future work aims to create a
cutting-edge AI-powered drowsiness detection system that is both inclusive and highly effective
in ensuring road safety for all. By training multiple models we aim to provide the best model
as a proof of concept that these biases can be mitigated in drowsiness detection.

3 Use Case Requirements
• REQ-01: Test the generated models offline on a diverse dataset containing different racial
profiles and facial characteristics that will ensure that our model is unbiased and fair and
doesn’t lose performance.
• REQ-02: Train a generic model that can be fine-tuned and be customization. We plan on
using VisionTransformer models as our basic architecture since they showed promising
results on training on different generic tasks. This can ensure we have a base model and
it can be fine-tuned to the user specific expression. That might make the model better
on individual performance and it won’t need a lot of data to be fine-tuned on the user
expression.
• REQ-03: The model should be able to make predictions real-time on a mobile device.
We want to use a mobile device which is widely available as our capture medium for the
user’s data. Our model should be able to make predictions in real-time when running on
a mobile device.
• REQ-04: On top of our model being unbiased in terms of race and facial characteristics it
will also need to be independent of factors such as the use of sunglasses. Also it should
be independent of different weather and lighting settings.

701