Globally, traffic accidents claim approximately 1.19 million lives each year according to WHO, with pedestrians, and other vulnarable road users, accounting for more than half of these fatalaties. For many years, Sweden has had a zero-vision policy regarding traffic fatalities, but despite this, around 200 people die in traffic each year, according to figures from the government agency Transport analysis.
Improving the prediction of pedestrian behaviour could play an important part in reducing these numbers. In a recent study, Chi Zhang and colleagues use machine learning to predict the behaviour of pedestrians in various traffic situations. They have been able to both predict pedestrian's decisions to cross a road with better accuracy than before, and more quickly predict pedestrian trajectory.
Intentions predicted with better accuracy
The researchers used machine learning to predict pedestrians' intentions—specifically their decision to cross a road.
“In over 90 percent of cases, the study was able to successfully predict whether a pedestrian would cross the road or wait, which is 4 percent higher than existing models. With over 94 percent accuracy, we could successfully predict whether the pedestrian would use a zebra crossing, a result also 4 percent better than existing models”, says Chi Zhang.
One of the datasets used in the study was collected from participants in Japan and Germany. The study observed 60 pedestrians from Germany and 60 pedestrians from Japan as they interacted with simulated vehicles. Participants wore VR headsets and were tasked with navigating a simulated traffic scenario, moving from one location to another location on the other side of the road within one minute, while researchers recorded their movements and actions.
“Although the VR environment was designed to closely mimic real-world scenarios, participants were aware that the environment wasn’t real, which may have influenced their behavior to some extent. However, in most cases, we were able to successfully predict their intentions, and the comparison between the two countries suggests that it’s possible to draw general conclusions from the results”, says Chi Zhang.
Factors influencing the decision to cross a road
Chi Zhang’s dissertation also analyses how some key factors influence pedestrians' intentions to cross a road—an area that has not been extensively explored in previous studies.
"We found that the presence of crosswalks/zebra crossings, waiting time, walking speed, and missed crossing chances strongly influence pedestrian behaviour", says Chi Zhang.
"For example, a pedestrian who misses several “gaps” between cars in traffic becomes more likely to take risks. This was evident regardless of the country observed”, Chi Zhang adds.
However, the researchers did note some differences between countries. In the data collected in Japan, participants moved faster but chose larger gaps between cars at crossings. In Germany, participants walked more slowly but took greater risks at crossings.
Predicting pedestrian trajectory faster than before
To predict pedestrians trajectory, the researchers used deep learning and built upon existing models by adding a custom subnetwork containing information about pedestrians, such as how they interact and influence each other.
The researchers used real-world data from two different open source platforms—one from self-driving cars in the U.S. and another from fimed traffic situations in Switzerland and Cyprus.
The researchers' model used the data to analyse pedestrian's movement by using just a few seconds from a pedestrain's past trajectory, and from that information the model could predict the the pedestrian's next step.
“Previous studies have also been able to predict pedestrian trajectory, but our study delivers results much faster than before. This quick data processing has the potential to be used in the development of self-driving cars, where it is crucial for the vehicle to receive data as quickly as possible”, says Chi Zhang.
Chi Zhang believes the study’s findings on pedestrian behavior in traffic are valuable both for the development of autonomous vehicles and for urban traffic planning. She hopes the results will ultimately contribute to reducing fatal pedestrian accidents.
Text Natalija Sako
