Abstract

Road accidents are an important issue for our societies, responsible for millions of deaths and injuries every year representing a very high cost for society. In this thesis, we evaluate how machine learning can be used to estimate the risk of accidents in order to help address this issue.
Previous studies have shown that machine learning can be used to identify the times and areas of a road network with increased risk of road accidents using road characteristics, weather statistics, and date-based features. In the first part of this thesis, we evaluate whether more precise models estimating the risk for smaller areas can still reach interesting performances. We assemble several public datasets and build a relatively accurate model estimating the risk of accidents within an hour on a road segment defined by intersections.
In the second part, we evaluate whether data collected by vehicle sensors during driving can be used to estimate the risk of accidents of a driver. We explore two different approaches. With the first approach, we extract features from the time series and attempt to estimate the risk based on these features using classical algorithms. With the second approach, we design a neural network directly using the time series data to estimate the risk. After extensively tuning our models, we managed to reach encouraging performances on the validation set, however, the performances of our two models on the test set were disappointing. This led us to believe that this task might not be feasible, at least with the dataset used.