Today’s world is a vast maze connected by layers of concrete and asphalt that give us the luxury of navigating in a vehicle. With much of our road-related advances, GPS is allowing us to fire fewer neurons thanks to mapping applications. Cameras warn us of potentially expensive scuffs and scrapes, and autonomous electric cars have lower fuel costs; our security measures have not been updated. We continue to rely on a constant diet of traffic signs, trust, and the steel around us to get from point A to point B safely.
To anticipate the uncertainty associated with accidents, scientists from the Laboratory for Science at MIT Computing and Artificial Intelligence (CSAIL) and the Qatar Center for Artificial Intelligence have developed a deep learning model that predicts accident risk maps with very high resolution.
Data, road maps, satellite images and GPS traces describe hazard maps the expected number of accidents over a certain period of time in the future in order to identify high risk areas and predict future road accidents.
These types of hazard maps are typically taken at much lower resolutions, in the hundreds of meters range, which means that important details are missing when roads become blurry. However, these maps are 5×5 meter grid cells, and the higher resolution brings new clarity: Scientists have found that a motorway, for example, carries a higher risk than nearby service roads, and driveways that meet the motorway are even higher risk than other motorways.
By capturing the underlying risk distribution that determines the likelihood of future accidents at all locations, and without historical data, we can find safer routes and enable auto insurance companies to provide tailored insurance plans based on customer travel routes and city planners design safer roads and even predict future accidents, says Songtao He, a graduate student at MIT CSAIL and lead author of a new article on the research.
Although car accidents are rare, they cost around 3 percent of global GDP and are the leading killer of children and young adults. This scarcity makes deriving maps with such high resolution a difficult task. The average annual probability of an accident in a 5×5 grid cell is roughly one in 1000 – and they rarely occur twice in the same place, if there has previously been an accident in the vicinity.
The team approach uses a larger network to collect critical data. Identify high-risk locations using GPS track patterns that provide information on traffic density, speed, and direction, as well as satellite imagery that describes road structures such as the number of lanes, whether there is a hard shoulder or a lot of pedestrians. Even if no accidents have been recorded in a high-risk area, it can still be identified as a high-risk area based on its traffic pattern and topology alone.
To evaluate the model, the scientists used crashes and data from 2017 and 2018, and tested its performance to predict crashes in 2019 and 2020 and track it.
Our model can be generalized from one city to another by combining multiple traces from seemingly disjointed data sources. This is a step towards generalized AI as our model can predict accident maps in uncharted areas, said Amin Sadeghi, senior scientist from the Qatar Computing Research Institute (QCRI) and author of the article. The model can be used to derive a useful accident map, even without historical accident data, which could have a positive impact on urban planning and policy-making when comparing imaginary scenarios.
The data set covered 7,500 square kilometers of Los Angeles, New York City, Chicago, and Boston. Of the four cities, Los Angeles was the most unsafe with the highest accident density, followed by New York City, Chicago and Boston.
If people can use the Risk Map to identify high-risk stretches of road, they can take steps in advance to reduce the risk of their journeys. Applications like Waze and Apple Maps have tools that make incidents work. But we’re trying before the crashes happen, before they happen, he said.
He and Sadeghi co-wrote the article with Sanjay Chawla, QCRI research director, and MIT computer and electrical engineering professors Mohammad Alizadeh, Hari Balakrishnan and Sam Madden Vision.
