Researchers use artificial intelligence to predict the behavior of road users

For an autonomous vehicle to travel safely, it is essential to be able to predict the behavior of other road users. A CSAIL research team from MIT (Massachusetts Institute of Technology) has teamed up with researchers at the Institute for Interdisciplinary Information Sciences (IIIS) at Tsinghua University in Beijing to develop a new ML system that could one day drive driverless cars closer. Drivers, cyclists and pedestrians in real time. They titled their research: M2I: Interactive Forecast from Factorized Marginal Trajectory Forecast “.

Kiao Sun, Junru Gu, Hang Zhao were the IIIS members who participated in the study while Jin Huang and Brian Williams represented MIT.

People are unpredictable, which actually makes it very difficult to predict the behavior of road users in urban areas. The AI ​​solutions currently in use are very simple: for them, a pedestrian, for example, can stay on the same sidewalk without trying to cross. If they expect pedestrians to cross, the robot simply parks the car to avoid them, some only predict the movement of a single road user.

Share to make better predictions

Trajectory predictions are widely used by intelligent driving systems to predict future movements of nearby agents and to identify risky situations to enable safe driving. For the team, existing models are great for predicting the marginal trajectory of single agents, but do not provide an answer for traffic in urban areas where many users communicate, the prediction space increasing rapidly with their numbers.

MIT researchers have come up with a seemingly simple solution to this complex problem: they break down the problem of predicting a multi-agent behavior into several smaller parts and then attack each one individually so that a computer can solve this complex task in real time. They call this method M21. Their behavior prediction framework first estimates the relationship between the two road users: which car, cyclist or pedestrian has the right of way and which agent will grant the right of way … then uses these relationships to predict the future trajectory of multiple agents.

In a huge dataset compiled by the self-driving company Wemo, the routes estimated by the M21 have been shown to be more accurate than the actual ML models compared to other ML models. (The MIT strategy also surpassed the recently published model). Moreover, dividing the problem into sub-problems allows them to use less memory.

Jean “Cyrus” Huang, a graduate student in the Department of Aeronautics and Astronomy and a research assistant in Brian Williams’ lab, professor of aeronautics and astronomy, and head of the Computer Science and Artificial Intelligence Laboratory (CSAIL), said:

“It’s a very intuitive idea, but no one has fully explored it before, and it works pretty well. Simplicity is definitely a plus. Achieving performance এর has a lot of potential for the future.

M21 method

In this work, researchers explored the underlying relationship between interacting agents. The M21’s algorithm has two inputs: a map with the past trajectory of cars, cyclists and pedestrians interacting in a traffic environment such as a four-way intersection as well as the location of the road, lane configuration, etc.

Using this information, a relationship prediction assumes which of the two agents has the right to pass first, classifying one as passer and the other as yielder. Then, a prediction model, called a marginal prediction, assumes the trajectory of the passing agent, since this agent behaves independently.

A second prediction model, known as a conditional predictor, then guesses what the passing agent is going to do based on the actions of the passing agent. The system predicts a number of different trajectories for the dealer and setter, calculates the probability of each individually, and then selects six joint results with the highest probability of occurrence.

The M2I method provides a prediction of the trajectory of these agents for the next eight seconds. It can slow down a vehicle so that a pedestrian can cross the road, then increase the speed when they clear the intersection. In another instance, the car waited for several cars to leave before moving off the sidewalk to the busy main road.

Waymo Open Motion Dataset Test

Researchers have trained the models in the WeMo Open Motion Dataset, which contains millions of real-life traffic scenes involving vehicles, pedestrians and cyclists, recorded by sensors and leader (light detection and ranging) cameras. 8 mounted in the autonomous vehicle of the company They only captured scenes where several agents were involved.

They then compare the six predictive patterns of each method, weighed by their confidence level, with the actual trajectory followed by cars, cyclists and pedestrians in one scene. Their method was the most accurate. The M21 also surpassed models based on a metric known as the overlap rate; If two trajectories overlap, it indicates a collision. M2I had the lowest overlap rate.

Jin Huang says:

“Instead of just creating a complex model to solve this problem, we have adopted a method that is similar to the way a person thinks when arguing about interactions with others. A person does not argue about hundreds of combinations of future behaviors. Makes quick decisions. Another advantage of M2I is that it breaks down the problem into smaller pieces, making it easier for the user to decide on the model.

On the other hand, the framework cannot account for cases where two agents influence each other, such as when two vehicles are advancing at each four-way stop because drivers do not know who should yield. The team wants to address this limitation in future work. He hopes to use his method to mimic realistic interactions between road users, which will verify self-driving vehicle scheduling algorithms or generate large amounts of synthetic driving data to improve model performance.

Article Source: “M2I: Interactive Predictions from Factored Marginal Trajectory Predictions” Kiao Sun, Jin Huang, Junru Gu, Brian C. By Williams and Hang Zhao. March 28, 2022 Computer Science Robotics.
arXiv: 2202.11884

Leave a Comment