Predictive analysis for road accidents using a tree-based and deep learning fusion system

Abstract

While road accidents’ prediction has been of crucial importance in the development of intelligent transportation technologies; a profound analysis within the driver-vehicle-environment system is no doubt of great interest and necessity. Three categories of features namely vehicle kinematics, driver inputs and environmental conditions collected using a desktop driving simulator have been systematically recorded in order to outline a fusion strategy based on various base classifiers and a Meta classifier that learns from base classifiers’ results to acquire more efficient accidents’ predictions. Highly heuristic optimized tree-based models namely AdaBoost, XGBoost, RF along with the MLP deep learning technique have been endorsed to establish effective predictions. Furthermore, to ensure that the proposed system provide superior and stable decisions as road accidents are generally unexpected and occur rarely, an imbalance-learning approach was conducted to add to the current knowledge by adopting three performant balancing strategies: ROS, SMOTE and ADASYN. To the best our knowledge, there has been a limited interest at adopting a fusion-based system examining the impact of real-time features’ combinations and fused tree-based models along with deep learning technique as meta-classifier on the prediction of road accidents while taking into account class imbalance. The findings depict that the superior performance of the proposed fusion system with precision, recall and f1-score over 90%. As a whole, the results highlight the significance of the explanatory features related to potential accidents and can be employed in designing efficient intelligent transportation systems.