Anticipate the behavior of others on the road | MIT news

Humans may be one of the biggest barriers to completely banning self-driving vehicles from city streets.

If the robot is to drive a vehicle safely through downtown Boston, it must be able to predict what nearby drivers, cyclists and pedestrians will do next.

Predicting behavior is a difficult problem, however, and current AI solutions are either too simplistic (you might assume pedestrians always walk in a straight line), too conservative (to avoid pedestrians, the robot leaves the car in the parking lot), or it can predict Only the following moves for a single agent (Methods usually carry many users at once.)

MIT researchers have created a deceptively simple solution to this complex challenge. They break the multifactorial behavior prediction problem into smaller parts and deal with each one individually, so that the computer can solve this complex task in real time.

Their Behavior Prediction Framework first guesses the relationships between two road users—which car, cyclist, or pedestrian has the right of way, and which agent will yield—and uses these relationships to predict the future paths of various agents.

These estimated routes were more accurate than those in other machine learning models, compared to real traffic flows in a massive data set compiled by self-driving company Waymo. MIT’s technology even outperformed the recently published Waymo model. Because the researchers broke the problem down into simpler parts, their method used less memory.

“This is a very intuitive idea, but one that no one has fully explored before, and it works well. Simplicity is definitely a plus. We compare our model with other latest models in the field, including the one from Waymo, the industry leader, and ours achieves the highest performance in this challenging criterion.says co-lead author Shane “Cyrus” Huang, a graduate student in the Department of Aeronautics and Astronautics and a Research Assistant in the Laboratory of Brian Williams, Professor of Aeronautics and Astronautics and Professor of Aeronautics and Astronautics Member of the Computer Science and Artificial Intelligence Laboratory (CSAIL).

Huang and Williams were joined in the paper by three researchers from Tsinghua University in China: co-lead author Qiao Sun, research assistant; Junro Joe, graduate student; and senior author Hang Zhao PhD ’19, assistant professor. The research will be presented at the Computer Vision and Pattern Recognition Conference.

multiple small Models

The researchers’ machine learning method, called M2I, takes two inputs: past paths of cars, cyclists and pedestrians interacting in traffic settings such as a four-way intersection and a map of street locations, lane configurations, etc.

Using this information, the relationship predictor indicates which of the two workers has the right of way first, classifying one as a bystander and the other as a worker. Then the prediction model, known as the marginal predictor, guesses the trajectory of the passing factor, because that factor acts independently.

The second prediction model, known as the conditional predictor, then guesses what the returning agent will do based on the actions of the passing worker. The system predicts a number of different paths for gentlemen and bystanders, calculates the probability of each individually, and then selects the six combined outcomes with the highest probability of occurrence.

M2I produces a prediction of how these agents will move through traffic over the next eight seconds. In one example, their method caused the car to slow down so pedestrians could cross the street, then increase speed when evacuating the intersection. In another example, the car waited until several cars had passed before turning from a side street onto a busy main road.

While this initial research focuses on the interactions between two factors, M2I can infer relationships between many factors and then guess their pathways by correlating several marginal and conditional predictors.

This simulation shows how the system developed by the researchers can predict the future paths (highlighted with red lines) of blue vehicles in complex traffic situations involving other cars, cyclists and pedestrians.

real world dshredding tests

The researchers trained the models using the Waymo Open Motion Dataset, which contains millions of real-time traffic scenes involving vehicles, pedestrians and cyclists recorded by sensors, Lidar (light and range detection) cameras, and cameras installed on the company’s self-driving vehicles. They focused specifically on cases with multiple agents.

To determine accuracy, they compared the six prediction samples for each method, weighted by their confidence levels, to actual paths taken by cars, cyclists, and pedestrians in a scene. Their method was the most accurate. It also outperformed the base models on a scale known as the overlap rate; If two paths overlap, this indicates a collision. M2I had the lowest interference rate.

“Instead of just building a more complex model to solve this problem, we took an approach much like how humans think when they think about interactions with others. Humans don’t think about all hundreds of sets of future behaviors. We make our decisions very quickly,” Huang says.

Another advantage of M2I is that, since it breaks down the problem into smaller parts, it is easier for the user to understand the decision-making process of the model. In the long term, this could help users increase confidence in self-driving vehicles, Huang says.

But the framework cannot take into account situations where two factors mutually influence each other, such as when two vehicles are moving forward at a four-way stop because drivers are not sure who should give in.

They plan to address this limitation in future work. They also want to use their method to simulate real-world interactions between road users, which can be used to verify planning algorithms for self-driving cars or generate massive amounts of synthetic driving data to improve model performance.

“Forecasting the future trajectories of many interacting factors is not well explored and is very challenging to enable complete autonomy in complex scenes. M2I provides a very promising prediction method with relationship prediction to characterize marginally or conditionally predicted factors which greatly simplifies the problem,” wrote Masayoshi Tomizuka, Professor Sherrill and John Nearhout, Jr. Distinguished in Mechanical Engineering at UC Berkeley, WA. Zahn, research assistant professional, in an email. “The prediction model can capture the inherent relationship and interactions between agents to achieve cutting-edge performance.” The two colleagues were not involved in the research.

This research is supported in part by a Qualcomm Innovation Fellowship. The Toyota Research Institute also provided funds to support this work.

Leave a Reply

%d bloggers like this: