Crash Predicting Network with Graded Warning for Vehicle Collision Warning
Physical Sciences : Automotive
Available for non-exclusive licensing
- Risto Miikkulainen, Ph.D. , Computer Science
- Rini Sherony , Toyota Technical Center, USA, Inc.
- Kenneth Stanley, Ph.D. , Computer Science
- Nathaniel Kohl , Computer Science
Automobile accidents are the number-one cause of death for ages 3 to 33 in the United States today. Each driver has individual driving styles and habits that are difficult to predict, not to mention outside variables such as other vehicles, bad weather, and road conditions. Although it is extremely complex for artificial intelligence technology to make a judgment on the level of danger in each driving circumstance, such information is necessary in order to create an accurate collision predictor.
NeuroEvolution of Augmenting Topologies (NEAT) trains neural networks for sequential decision tasks. NEAT gains experience in predicting collisions using a simulator such as the Robot Auto Racing Simulator (RARS), which simulates vehicle dynamics (including skidding and traction) and interactions between multiple vehicles. The crash predictor uses NEAT’s recurrent network to analyze the possibility of a crash, including driving off the road and colliding with other cars. The algorithms works with range-finder and sonar inputs, as well as raw visual data, and provides a graded warning.
- Crash prediction prevents automobile accidents.
- Warnings are accurate and helpful to human drivers.
- Warnings can be customized to driver, car and conditions.
- Graded warning system helps judge danger level of a situation.
- By integrating over past states, can predict crashes that cannot be calculated only from a vehicle’s current state
- Trains both drivers and crash predictors
- Uses raw visual input from a simulated camera to predict the time to crash
Automotive industries and transportation safety industry may market product for consumers. Defensive driving and driver education schools may use the invention as a teaching tool.
- 1 U.S. patent issued: 7,565,231