Data-Driven Road Safety Evaluation

Predict road safety level with merely satellite map images!

AI for Good Foundation brought our attention to the traffic safety around the world. There are so many injuries and death due to vehicle accidents. The team wanted to figure out the relationship between traffic safety and road design. The findings on road design features can be used to advice civil engineers to build safer roads, and the road safety prediction technology can also be used to improve the driving safety for the autonomous vehicles.

The team worked on Convolution Neural Network to train the model from learning the satellite map. Firstly, the team focused on the Manhattan area because the roads there are crowded and there are ample open source data available.

In order to build the model, the team tackled down several difficult technical problems such as how to cut the map into smaller pieces, how to find the intersections in a satellite map and what hyperparameters should be modified. The team overcame the difficulty and obtain an outstanding accuracy in predicting road safety level of around 75%.