Autonomous driving, a prophetic puzzle

Driving is a subtle art of divination. or approx. In addition to respecting highway codes, a large proportion of decisions are based on anticipating what other road users are going to do, but also on their behavior in relation to one another. Pedestrian, cyclist or oncoming car, who will give way to others, who will enter first, who will give way to whom and how will others react?

This is exactly what autonomous car control systems struggle to do. And this is exactly what an algorithm developed by a team from the artificial intelligence laboratory of the Massachusetts Institute of Technology (MIT) in the United States, involving three Chinese researchers from Tsinghua University, manages to do.

Interactive Prediction

Called M2I, this “interactive prediction” project will be presented in June 2022 at CVPR, an annual cycle of conferences on computer vision and pattern recognition. It applies to a variety of contexts, such as intersections controlled by traffic lights, access ramps to expressways, change of direction of various vehicles, interruption of cyclists, change of lanes, pedestrians crossing the road in the middle of traffic…

The algorithm was trained and tested by Google’s self-driving car division, Waymo, using a database of road scenarios drawn from real-world situations. Theory involves dividing a complex problem into several simpler ones. Simply put, the autonomous vehicle detects two road users and estimates the trajectory that each will describe. Based on which it determines two roles: one of the users will be the influencer, the other the one who will react. Clearly, the algorithm determines which agent will block its behavior on the other, then calculates the trajectory it follows within eight seconds of each.

See also  NASA's 'wonderful' discovery of Mars supported 'rich life' on the surface of the Red Planet. Science | News

Road scenarios are analyzed and processed by the MI2 algorithm. credit: Seisel MIT

reduce computation time

In this case, the algorithm extracts six possible trajectories, each of which is assigned a confidence level. Thus the program moves from one interaction to the next, not by analyzing the entirety of a scene, but by the evolution of traffic continuously and in real time. The process eventually reaches human reflexes while reducing computation time, which allows the autonomous car’s system to retain information, therefore, to make decisions more quickly.

However, this solution is not perfect. Some trajectory predictions, once they are compared to the actual trajectory of the database model, result in (virtual) collisions. But the M2I records the lowest collision rate ever compared to existing technologies.

You May Also Like

About the Author: Abbott Hopkins

Analyst. Amateur problem solver. Wannabe internet expert. Coffee geek. Tv guru. Award-winning communicator. Food nerd.

Leave a Reply

Your email address will not be published.