AI Researchers Simulate Peripheral Vision in Machine Learning Models to Improve Vehicle Safety
AI that can model peripheral vision could be used to develop new vehicle safety features. Photo by Christoph Reichwein/picture alliance via Getty Images.
Key Takeaways
- Computer vision researchers attempted to teach AI models to see in the periphery.
- Novel peripheral vision models could potentially be used to develop driver safety systems.
- Researchers hope to develop AI models that can accurately predict how humans detect objects at the edge of their visual field.
Although self-driving vehicles are the most obvious automotive use case for computer vision, the AI technology has many other potential applications in the field.
For instance, a recent study that used Machine Learning (ML) to simulate peripheral vision could lead to increased driver safety by helping AI models “see” more like humans.
Computer Vision Researchers Explore the Periphery
In the context of AI models that can indiscriminately process a continuous stream of visual information to scan for patterns, teaching neural networks the art of peripheral vision seems counterintuitive at first glance.
For humans, peripheral vision extends the visual field outside the area of focus. But if AI can process whatever visual data is thrown at it, what’s the point of focusing on a limited region of space?
As the recent study’s authors emphasized, however, if computer vision models can be taught to mimic human visual processing, they could be used to predict whether people will see objects they aren’t focused on.
Potential Safety Applications
One proposed application for peripheral vision models is to incorporate them into AI systems that alert drivers to hazards they might not see.
The researchers said their eventual goal is to develop a model that can predict human performance in the visual periphery with a high degree of accuracy. However, their initial attempts fell significantly short of human’s ability to perceive things out of the corner of their eye.
Human Application of AI
When processing visual data up to 10° away from the center of focus, the researchers found that AI and human image detection decreased at a similar rate. But after that, the ability of neural networks to recognize distinct objects dropped off a cliff, while human vision declined gradually to the edge of the visual field.
Understanding how humans are able to identify objects with limited visual information could help increase the performance of future AI models, especially in the low visibility conditions that are most dangerous for drivers.
