Deep Learning Helps Predict Traffic Accidents Before They Happen | MIT News


Today’s world is a great labyrinth, connected by layers of concrete and asphalt that offer us the luxury of vehicle navigation. For many of our road-related advancements – GPS lets us trigger fewer neurons through mapping applications, cameras alert us to potentially expensive scuffs and scratches, and self-driving electric cars have lower fuel costs – our security measures haven’t quite caught up. We always rely on a constant regime of traffic lights, trust and the steel around us to get us safely from point A to point B.

To stay ahead of the uncertainty inherent in accidents, scientists at MIT’s Computer and Artificial Intelligence Laboratory (CSAIL) and the Qatar Center for Artificial Intelligence have developed a deep learning model that predicts risk maps of crash at very high resolution. Fueled from a combination of historical accident data, road maps, satellite imagery and GPS tracks, hazard maps describe the expected number of accidents over a future period of time, in order to identify the high-risk areas and predict future accidents.

Typically, these types of hazard maps are captured at much lower resolutions that hover around hundreds of meters, meaning crucial details have to be ignored as the roads together become blurry. These maps, however, are 5×5 meter grid cells, and the higher resolution brings new clarity: Scientists have found that a highway road, for example, poses a higher risk than neighboring residential roads, and that the ramps that merge and exit the freeway have an even higher risk. risk than other roads.

“By capturing the underlying risk distribution that determines the likelihood of future accidents at all locations, and without any historical data, we can find safer routes, allow auto insurance companies to offer insurance plans. personalized based on customer driving trajectories, helping city planners design safer roads and even predict future crashes, ”said Songtao He, doctoral student at MIT CSAIL, lead author of a new research article.

Although car accidents are rare, they cost around 3% of global GDP and are the leading cause of death in children and young adults. This rarity makes the deduction of maps at such a high resolution a delicate task. Accidents at this level are not widely dispersed – the average annual probability of an accident in a 5×5 grid cell is about one in 1,000 – and they rarely occur twice at the same location. Previous attempts to predict accident risk have been largely “historical”, as an area would only be considered high risk if there had been an accident nearby.

The team’s approach casts a wider net to capture critical data. It identifies high-risk locations using GPS track models, which provide information on the density, speed and direction of traffic, and satellite imagery describing road structures, such as the number of lanes, s ‘there is a shoulder or if there are a large number of them. pedestrians. Then, even if a high risk area has not recorded any accidents, it can still be identified as high risk, based only on its traffic patterns and topology.

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. Many places were identified as high risk, although they didn’t had recorded no accidents, and also experienced accidents during the years of follow-up.

“Our model can be generalized from city to city by combining multiple clues from seemingly unrelated data sources. This is a step towards general AI, because our model can predict crash maps in uncharted territories, ”explains Amin Sadeghi, senior scientist at the Qatar Computing Research Institute (QCRI) and author of the article. “The model can be used to derive a useful crash map even in the absence of historical crash data, which could translate into positive use for city planning and policy making by comparing imaginary scenarios. ”

The data set covered 7,500 square kilometers of Los Angeles, New York, Chicago and Boston. Of the four cities, LA was the most dangerous, as it had the highest accident density, followed by New York, Chicago, and Boston.

“If people can use the risk map to identify potentially high-risk road segments, they can take action in advance to reduce the risk of the trips they make. Apps like Waze and Apple Maps have crash functionality tools, but we try to anticipate crashes – before they happen, ”He says.

He and Sadeghi wrote the article alongside Sanjay Chawla, research director at QCRI, and MIT electrical engineering and computer professors Mohammad Alizadeh, Hari Balakrishnan and Sam Madden. They will present the article at the 2021 International Computer Vision Conference.


Source link

Previous DVIDS - News - Change of command in Michigan's cyber community
Next Altaba to Allocate $ 400 Million to Verizon for Yahoo Settlement Appeal

No Comment

Leave a reply

Your email address will not be published. Required fields are marked *