Page 735 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

2 Use Case Description

2�1 Description

Autonomous driving represents the inevitable trajectory of the automotive industry. However, 4.9: Accessibility
the widespread adoption of fully autonomous technologies is not an immediate objective,
nor will it occur uniformly across different regions. Consequently, existing Advanced Driver-
assistance Systems (ADAS), such as Driver Monitoring Systems (DMS) technologies, remain
highly relevant as they can anticipate and mitigate potential accident risks [1].

Driver drowsiness is a major contributor to road accidents globally, posing serious risks to both
drivers and the public. According to the European Commission, fatigue-related crashes account
for up to 20% of all road traffic accidents in Europe. In the United States, the National Highway
Traffic Safety Administration (NHTSA) estimates that drowsy driving causes approximately
100,000 police-reported crashes annually, resulting in over 1,500 deaths and 71,000 injuries.
One widely cited real-world incident occurred in 2014, when a Walmart truck driver—allegedly
awake for more than 24 hours—caused a fatal crash involving comedian Tracy Morgan, bringing
national attention to the dangers of fatigued driving [2].

In response to this critical safety issue, Driver Monitoring Systems (DMS) powered by AI have
emerged as vital tools to detect drowsiness and prevent accidents. However, current systems
often fall short in reliably identifying drowsiness across diverse populations due to biased
training data and overly generalized models.

Upon evaluating a DMS system utilizing state-of-the-art facial landmark recognition models, it
was observed that these models exhibit limitations in accurately identifying facial landmarks
across individuals of diverse racial backgrounds and varying facial characteristics, such as the
presence of facial hair [2-3]. This disparity may result in false negative outcomes in the drowsiness
detection alerts, thereby compromising its reliability and effectiveness. The underlying causes
of these limitations may include insufficient racial diversity in the training datasets and the
inherent simplicity of the models, which may prioritize generalized facial features over nuanced
individual variations. Addressing these challenges is essential to ensuring the inclusivity,
accuracy, and equitable performance of DMS technologies across diverse populations.

In line with the principles of inclusivity, fairness, and the ethical deployment of Artificial
Intelligence, we suggest a use case with native alternative data models designed specifically
to address the identified limitations. These models have a strong emphasis on ethical
considerations and equitable performance across diverse populations, ensuring that individuals
of all backgrounds are accurately represented. Importantly, this commitment to native fairness
and inclusivity will be pursued without compromising the overall efficacy and reliability of the
model.

To evaluate the fairness and unbiasedness of our model we plan to use first of all some
automated tools. Upon our initial research we encountered several automated tools with the
most promising one being Fairlearn [7]. Fairlearn can be integrated seamlessly in our training
pipeline and give us metrics on the predictions of the model such as demographic parity,
predictive parity and calibration curves just to name a few key metrics to assess the fairness
of a model [8]. We also plan to deploy automatic pipelines to check the bias of a dataset by
identifying unbalanced in any dataset between the prediction classes and also creating a

699