An Ensemble of Multiple Boosting Methods Based on Classifier-Specific Soft Voting for Intelligent Vehicle Crash Injury Severity Prediction

Abstract

To mitigate the crash injury severity of intelligent vehicles when faced with an emergency, it is critical to build an applicable crash injury severity model. An ensemble model (CSSV-AGX) of AdaBoost (Adaptive Boosting), GBDT (Gradient Boosting Decision Tree), and XGBoost (eXtreme Gradient Boosting) based on Classifier-specific Soft Voting was proposed in this paper. Several key factors were analyzed to find out relationships between accident factors (for instance, speed) and distribution of different injury severity of occupants via our CSSV-AGX model. When faced with an emergency, our analysis will help intelligent vehicles to make appropriate decisions and therefore mitigate occupants' injury severity. A total of 28071 samples released by National Automotive Sampling System - Crashworthiness Data System (NASS-CDS) were used in this study. Experimental results demonstrate that our CSSV-AGX model (50.96%) is superior to traditional machine learning methods (Support Vector Machine (49.05%), Multi-layer Perceptron (48.22%), and Random Forest (49.96%)) in multi-classification problem by accuracy. By sensitivity analysis, we concluded that the vehicle speed, weight, and occupants' age, which are key factors to intelligent vehicles, have a great impact on crash injury severity.