Deep learning helps predict traffic accidents before they happen

Today's world is one gigantic labyrinth connected by layers of concrete and asphalt that allows us to travel by car. In many aspects of road progress, GPS allows fewer neurons to be used thanks to mapping applications, cameras warn of potentially costly scratches and scratches, and electric autonomous vehicles are cheaper to refuel. .

Security measures were not applied properly. comfort. We rely on traffic lights, trust, and a relentless diet of steel to get us safely from A to B.

To avoid the uncertainty associated with accidents, scientists at the Massachusetts Institute of Technology's Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Qatar Center for Artificial Intelligence have developed a deep learning model that predicts accident risk maps at very high resolution.

A hazard map based on a combination of historical crash data, road maps, satellite imagery and GPS footprints identifies areas of high risk and describes the expected number of crashes over a specified period in the future to  predict future accidents.

Typically, these types of hazard maps are recorded at a much lower resolution - hundreds of meters - so when the road is blurred, important details  are hidden. However, this map is a 5x5 meter grid cell, and the higher resolution provides new clarity.

Scientists have found that car roads, for example, pose a higher risk than nearby residential roads. much higher risk. It is more dangerous than other roads. “Capturing baseline risk distributions that determine the likelihood of future accidents at any location, finding safer routes without  historical data, enabling auto insurers to deliver customized insurance plans based on customer trajectories, and helping city planners design safer roads.

MIT CSAIL PhD student Heo Songtao, lead author of a new paper on this study, says. Although rare, automobile accidents account for about 3% of global GDP and are the leading cause of death among children and adolescents.

This scarcity makes it difficult to render maps at  high resolutions. Failures at this level are sparsely scattered. The average annual probability of failure for a 5x5 grid cell is about 1 in 1000, and  rarely occurs twice in the same location. Previous attempts to predict accident risk have been mostly "historical". This is because the area is considered high risk only  if  a previous collision has already occurred nearby.

The team's approach extends the network to collect mission-critical data. High-risk locations are identified using GPS trajectory templates that provide information about density, speed, and direction of travel, and satellite imagery that describes road structures such as  number of lanes, shoulders, or many pedestrians. . .

To evaluate the model, the scientists used defects and data from 2017 and 2018 and tested their predictive performance of 2019 and 2020 outages. Many locations were identified as high risk, but no outages were reported and outages occurred during follow-up. age. “Our model can generalize from one city to another by combining multiple clues from seemingly unrelated data sources. This is a step towards general AI as our model can predict thought maps in uncharted territory,” said Amin Sadeghi, senior researcher and article author at the Qatar Computing Research Institute (QCRI).

“This model can be used to generate useful incident maps even in the absence of historical incident data, and can be actively used in urban planning and policy making by comparing what-if scenarios.” The data set covered 7,500 square kilometers in Los Angeles, New York, Chicago, and Boston. Of the four cities, Los Angeles was the least secure with the highest crash density, followed by New York, Chicago and Boston.

“If people can use hazard maps to identify potentially hazardous parts of the road, they can take early action  to reduce the risks of moving journeys. Apps like Waze and Apple Maps have tools to handle incidents, but we're trying to be prepared for outages before they happen,” he says.

He and Sadegi co-authored the article with Sanjay Chawla, research director at QCRI, and Mohammad Alizadeh, Hari Balakrishnan, and Sam Madden, MIT professors of electrical  and computer science. They will present their papers at the 2021 International Conference on Computer Vision.