MIT and QCRI Deep Learning Predicts Traffic Accidents

And forecasting generalizes to multiple cities. 

Deep Learning Helps Predict Traffic Crashes Before They Happen

MIT News, Rachel Gordon, October 12, 2021

A deep learning model trained on historical traffic crash data, road maps, satellite imagery, and global positioning system trajectory patterns can generate high-resolution crash risk maps. Scientists at the Massachusetts Institute of Technology and the Qatar Computing Research Institute (QCRI) developed the model, which yields risk maps that can define the expected number of crashes over a future period, identifying high-risk areas and forecasting future collisions. The maps are composed of 5x5-meter grid cells, a resolution that shows highway roads, for example, have a greater risk for traffic accidents than nearby residential roads, while highway ramps have higher risk than other roads. QCRI's Amin Sadeghi said, "Our model can generalize from one city to another by combining multiple clues from seemingly unrelated data sources."