The world is a prominent, confusing place. We have all the technology to make navigation more accessible. However, we still rely on traffic lights and steel for safety measures because there’s no telling when something will go wrong in this maze-like landscape of concrete & asphalt roads that afford us luxury while also being connected by an interconnected webbing of wires overhead making sure nothing gets left behind from point A – B.
In the United States, traffic accidents cost about 3% of our country’s GDP and are a leading cause of death among children as well. Accident risk maps can help monitor drivers for patterns that might indicate an increased likelihood at risk in their area or type of vehicle involved with crashes to determine if any action needs to be taken accordingly.
The MIT CSAIL and Qatar Center for Artificial Intelligence developed a deep learning model that can predict the number of crashes over time in areas with high risk. From historical data, road maps, satellite imagery, and GPS traces, they created very accurate crash prediction models using an artificial neural network system called “random forest,” which is fed on historical information and future projections about traffic conditions along roads where accidents might happen.
Maps of risk are typically captured at lower resolutions that blur together the lines between roads because they’re so detailed. Maps made up 5×5 meter grid cells, however – with higher resolution than normal- bring new clarity: Scientists found highways have a much greater chance than residential streets and ramps merging or exiting from one highway has an even higher chance in comparison to other road types.
The odds of a crash in 5×5 grid cells are about one-in-1,000 — but they’re not as high anywhere else on the map. The old method for predicting these risks was “historical” because it only looked at crashes that had happened nearby before; areas were considered risky if there was another incident nearby.
The proposed deep learning model can identify high-risk locations using GPS trajectory patterns, which give information about density, speed and direction of traffic. It also identifies places with no crashes or few recorded ones as “high risk” due to their topology alone.
The scientists were able to make predictions about crashes in locations where there had been none before. They did this by using data from 2017 and 2018, with Predictions made for 2019-2020 proving accurate as well.
The dataset used in this research covered 7500 square kilometers from Los Angeles, New York City, Boston, and Chicago. LA had the highest crash density of these four cities, followed by New York, then Chicago and Boston.