How a driver interacts with a car while driving — where they are looking, where they are focused and where their hands are — has a big impact on safety. However, to evaluate a driver’s head pose, gaze and hand position while driving, as well as the car’s lane position, used to require researchers to manually view, analyze and record hours of data.
Could we create automated and shareable software tools to detect and track a driver’s movements in natural driving environments?
A test vehicle was outfitted with 8 cameras facing both inside and outside. Inside, the cameras were positioned to view the driver’s face and hands, as well as the foot pedal and cabin. Outside cameras recorded lane departures.
We recorded and observed hand activity between the steering wheel, instrument panel and shifter, to better understand how the driver interacts together with all three regions.
We recorded and calculated the driver’s head pose using the facial landmarks and their corresponding location on a 3D generic face model.
A new lane detection algorithm called LASeR (Lane Analysis using SElective Regions) was used to analyze driving data for lane change events.
This detection algorithm performed greater than 80% accuracy with 5 second observation window at less than 1 in 1000 false positives.
By automating the data capture and analysis, we are able to have quicker, more thorough and more cost-effective access to data regarding driving behavior. By having a better understanding of this behavior, we can help inform in-vehicle technology decisions, safety features and driver education.
In other words, the faster we can analyze, the faster innovation can happen.
Historically, automobile restraints have been designed for, and tested with, average adult sizes, shapes, and driving postures. But as drivers come in a variety of shapes and sizes, they also have a variety of safety and comfort needs.
Can we create virtual models that more accurately reflect real world drivers?
In order to determine how these ATDs compare to today’s real world drivers, we recruited 200 volunteers, of which two thirds were over 60 years of age.
What we learned was that current ATDs may not represent how the seat belt fits certain individuals, particularly those who are larger or who may not wear their seat belt optimally.
By scanning and studying how people sit and use seat seat belts, we designed pose-able body shape models to represent variations in age, body mass, gender and posture. These models will be shared with the auto industry.
Pedestrian Pre-Collision Systems (PCS) are designed to help prevent or mitigate collisions. Some automakers have incorporated this technology, which often uses a combination of radar, lidar and/or cameras to help detect objects in front of a vehicle.
But as of now, there is no standard process or equipment used across the automotive industry to evaluate the performance of each vehicle’s Pedestrian PCS.
Could we help develop tests and tools based on real world scenarios, to create a standard evaluation methodology for Pedestrian PCS to share with the automotive industry?
We developed a new model for a fully-articulated pedestrian mannequin, with arms and legs that move like a walking human, for use in testing pedestrian PCS systems. We called this fully-articulated mannequin “Steve”.
By radar scanning 9 different body types, we were able to create a material system for Steve’s skin (a combination of metal and fabric) that allows Pedestrian PCS systems to detect him as if he were human.
To help match test conditions to real-world driving environments, we gathered naturalistic driving data on pedestrian-vehicle interactions. To gather this data, 110 vehicles were equipped with data recorders that collected 93TB of driving data over 1.44 million miles.
Steve was put into a variety of different scenarios to mimic real life conditions:
With Steve, we created a groundbreaking, fully-articulated mannequin that walks and shares the same radar cross-section as a human. Steve is being shared within the industry to help standardize the testing of Pedestrian PCS in testing scenarios designed to mimic the real world.