Artificial intelligence researchers have developed a deep learning model to predict traffic accidents before they happen

Source: https://news.mit.edu/2021/deep-learning-helps-predict-traffic-crashes-1012


The world is a prominent and 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 is no telling when something will go wrong in this labyrinthine landscape of concrete and asphalt roads that afford us luxury while being connected by an interconnected strap of wires overhead ensuring nothing is overlooked from point A – B.

In the United States, traffic accidents cost approximately 3% of our country’s GDP and are also a leading cause of death in children. Crash risk maps can help monitor drivers for patterns that could indicate an increased likelihood of risk in their area or the type of vehicle involved in crashes to determine if action should be taken.

the MIT CSAIL and Qatar Artificial Intelligence Center developed a deep learning model that can predict the number of accidents over time in high-risk areas. Using historical data, road maps, satellite images and GPS tracks, they created highly accurate crash prediction models using an artificial neural network system called “random forest”, which is powered by historical information and future projections about traffic conditions along roads where accidents could occur. happen.

Risk maps are typically captured at lower resolutions which blur the lines between routes because they are so detailed. However, maps made up of 5 × 5 meter grid cells – with higher resolution than normal – bring new clarity: scientists have found that freeways are much more likely than residential streets and off-ramps. merging or exiting a freeway have even more chances compared to others. road types.

Source: https://openaccess.thecvf.com/content/ICCV2021/papers/He_Inferring_High-Resolution_Traffic_Accident_Risk_Maps_Based_on_Satellite_Imagery_ICCV_2021_paper.pdf

The odds of a crash in the 5×5 grid cells are around one in 1,000, but not as high elsewhere on the map. The old method for predicting these risks was “historical” because it only took into account accidents that had occurred nearby before; the areas were considered at risk if there was another incident nearby.

The proposed deep learning model can identify high-risk locations using GPS trajectory models, which give information about traffic density, speed and direction. It also identifies locations with no or few crashes recorded as “high risk” due to their topology alone.

Scientists were able to make predictions about crashes in places where there had been none before. To do this, they used data from 2017 and 2018, with the predictions made for 2019-2020 also proving to be accurate.

The dataset used in this research covered 7500 square kilometers of Los Angeles, New York, Boston and Chicago. LA had the highest accident density of these four cities, followed by New York, then Chicago and Boston.

Article: https://openaccess.thecvf.com/content/ICCV2021/papers/He_Inferring_High-Resolution_Traffic_Accident_Risk_Maps_Based_on_Satellite_Imagery_ICCV_2021_paper.pdf

Source: https://news.mit.edu/2021/deep-learning-helps-predict-traffic-crashes-1012

James G. Williams