Estimativa do Número de Vítimas Mortais a 30 Dias em Acidentes Rodoviários
João Moura Pires (superv.), Universidade Nova de Lisboa, Março 2015.
Keywords: Road Traffic Accidents, Visual Analytics, Black Spots
Abstract: By 2030, it is expected that road traffic accidents will become the fifth leading cause of death, when in 2013, it was placed as the eighth. In Portugal, in 2012, there were about 30 thousand traffic accidents with victims, which resulted in 718 fatalities at 30 days (human consequences within 30 days after the road traffic accident). Road traffic accidents have a big economic and social impact, therefore it is crucial to find ways of preventing them. However, the information about the number of fatalities at 30 days (a fundamental indicator used internationally) takes between 6 to 8 months to the acknowledged by the National Road Safety Authority. This causes a considerable delay in its analysis. Only after 2010, the real number of fatalities at 30 days started to be registered – before 2010, the number of fatalities was multiplied by 1,14. This correction factor presented a yearly error of about -10%.
In this work, it was tried to create an estimator of the number of fatalities at 30 days. In the first attempt, it was attempted to classify each accident so it could be understood if it was an accident were there would be divergences between the number of victims and the number of victims at 30 days, that is, if it is an accident where there were victims with injury gravity alteration. Regarding classification models, there were used the Support Vector Machine (SVM) and Classification and Regression Tree (CART). Since this is a problem of rare event detection – on a scale of 1:180, the models learn but don’t generalize.
In the second approach, it was tried to globally estimate, and for time periods superior to 1 month, the number of fatalities at 30 days. There were used a lot of regression techniques as well as fixed and sliding window for training. The solution, for 2012, obtained 6% of average monthly errors and 3% of annual error and, for 2013, it obtained 5.7% of average monthly errors and -1% of annual error. These results, when compared with the corretion factor of fatalities at 30 days, represent an improvement for the annual error of 65% and 90% for the years 2012 and 2013, respectively.