Traffic accidents impose significant problems in our daily life due to the huge social, environmental, and economic expenses associated with them. The rapid development in data science, geographic data collection, and processing methods encourage researchers to evaluate, delineate traffic accident hotspots, and to effectively predict and estimate traffic accidents.
In this study, Kaggle traffic accidents dataset that covers United Kingdom for the time period between 2012-2014 is investigated. Our methodology consists of three main techniques. First, Morans I method of spatial autocorrelation, and Getis-Ord Gi* statistics will be used to examine and relate traffic accidents dataset in terms of spatial and temporal features. Second, weighted features will be used as inputs for Deep Feed Forward Neural Network (DFFNN). Finally, the performance of the proposed DFFNN will be evaluated based on its accuracy, misclassification rate, precision, prevalence, histogram of errors, and confusion matrix.
Spatio-temporal analysis of the traffic accidents was per-formed in order to predict the severity degree of the accidents during 2012, 2013, and 2014. Using XGBoost and ETC, two hotspot maps were produced for each year. These maps were generated by taking into consideration the features with high weights.
For the accuracy assessment, the confusion matrix for XGBoost and ETC methods have been calculated by comparing the predicted severity with ground truth severity. XGBoost shows higher accuracy for all years as opposed to ETC. Fig. 3 shows the resulting maps for each year for the XGBoost method.
6 month (Feb, 2019 – Aug, 2019)
- Eng. Diena Aldogom
- Eng. Nour Aburaed
- Eng. Mina Talal Ahmed