Deep learning

As the world is huge and is connected by layers allowing us to easily navigate through its ways. Many of us have been habituated to GPS that can guide us through the roads. This is all possible with the help of map apps and camera alerts and reaches to our destinations. To jump over the uncertainties related to the maps and traffic, the Computer Science and Artificial Intelligence Laboratory otherwise called CSAIL of MIT along with the Qatar Center for Artificial Intelligence has come up with a deep learning model that can anticipate high-resolution crash risk maps. 

This deep learning model was developed by feeding on the combination of road maps, satellite imagery, GPS traces, historical data that can describe the risk maps of the expected number of crashes over a period of time in the future, to find out the high-risk areas and anticipate the future crashes. 

The lead author on a new paper about the research at MIT CSAIL Ph.D. student Songtao He, says that capturing the underlying risk distribution in the deep learning model that determines the probability of future crashes at all places, and without any historical data, enable auto insurance companies to provide customized insurance plans based on driving trajectories of customers. 

Even though car crashes are rare, they cost about 3 percent of the world’s GDP. The team approach casts a broad next to gain critical data. It sorts the high-risk locations using GPS trajectory patterns that give out information about the speed, density, and direction of the traffic. With the help of satellite imagery that describes the structures of the road such as lanes, or if there are pedestrians nearby, etc. If the high-risk area has no recorded crashes, it can still be identified as high-risk based upon the topology and traffic patterns. 

To find the right solutions, the researchers utilized the data of crashes from 2017 and 2018 and tested its performance at anticipating crashes in 2019 and 2020. Most of the locations were detected and identified as high risk even though there were no crashes and have been witnessed crashes the following year. 

The lead scientist at Qatar  Computing Research Institute or QCRI and also the author of the paper, Amin Sadeghi, says that the deep learning model can generalize from one city to another by adding multiple clues from unrelated data sources. With the help of AI, he says that the deep learning model can identify and predict the crash maps in uncharted territories. This dataset has included 7,500 kilometers from Los Angeles, Chicago, New York City, and Boston. When considered among the four cities Los Angeles was identified as the highest crash density after which follows New York City, 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 trips they take. Apps like Waze and Apple Maps have incident feature tools but we’re trying to get ahead of the crashes before they happen”, adds Sadeghi.