Steering Clear of Distractions

One of the biggest threats on the roadways is distracted driving. It is responsible for nearly half a million injuries per year in the United States alone, according to the Centers for Disease Control and Prevention. And the problem only seems to be getting worse as time goes by. The digital devices that now consume so much of our lives are constantly demanding our attention, and sometimes the urge to check that latest notification wins out over our common sense that is telling us to keep our eyes on the road.

Marketing campaigns and reminders in our phone apps do not seem to be having much impact on the problem, so more effective solutions are urgently needed. A team led by engineers at New Mexico State University believes that the solution to our technology-related distracted driving may actually be more technology. They have developed a prototype system that uses cutting-edge computer vision techniques to detect instances of distracted driving in real-time, which can then be used to trigger alarms or take other actions to get the driver reengaged with the task at hand.

In order to detect specific behaviors associated with distracted driving, the team chose to use an object detection algorithm. In many cases, these algorithms are very computationally expensive, which prevents them from running in real-time on the sort of hardware that might be found in a vehicle. Everything happens fast at highway speeds, making delayed results of this sort unacceptable. At the other end of the spectrum are very fast models that can operate in real-time on resource-constrained hardware, but they tend to lack accuracy. Such models might miss important cues, so they are also unsuitable for real-world use.

The team overcame these issues with existing systems by utilizing the Pretrained-YOLOv8 (P-YOLOv8) model. P-YOLOv8 is optimized for both speed and accuracy, overcoming the computational constraints and latency challenges that are common with traditional object detection models.

The researchers utilized the Distracted Driver Detection dataset from State Farm, which is comprised of 22,424 images categorized into ten distinct behaviors related to distracted driving to fine-tune their model. After doing so, they evaluated P-YOLOv8's capabilities in image classification and object detection, comparing its performance against established deep learning models like VGG16, VGG19, and ResNet. Accuracy levels proved to compare favorably with these models, yet the model size was reduced significantly.

While many traditional models either struggle with low accuracy or achieve high accuracy at the expense of slow detection speeds and high computational costs, P-YOLOv8 managed to strike a balance. It offers competitive accuracy — achieving an impressive 99.46 percent — while maintaining a lightweight model size of only 2.84 MB and a modest parameter count of 1.45 million.

The team’s work demonstrates that P-YOLOv8 may be a viable solution for building a distracted driver detection system that can operate in real-time on hardware that is practical for use in a real-world vehicle. Perhaps in the future systems like this one will help to save lives and prevent injuries by being combined with effective alert mechanisms.