Identifying High Risk Driving Scenarios Utilizing a CNN-LSTM Analysis Approach*

Abstract

High risk driving scenarios are critical for the deployment of highly automated vehicles virtual test. In this study, we have proposed a deep learning method to identify high risk scenarios from the field operation test (FOT) data. The proposed method tries to overcome the shortcomings of existing relevant studies for their limited utilizations of video data and mainly based upon instant kinematic indicators, which has led to high false alarm rate issue. In this study, a combined video analysis method (Convolutional Neural Network, CNN) and temporal feature analysis model (Long Short-Term Memory, LSTM) was proposed. To be specific, we used CNN-LSTM and Convolutional Neural Networks and Long Short-Term Memory (Resnet-LSTM) to perform the classifications for high risk scenarios and non-conflict scenarios. The empirical analyses have been conducted using commercial vehicle FOT data. And the results showed that the overall model performance (AUC index) in the test set could reach 0.91 with 83% accuracy rate. Finally, the future works have been discussed from the aspects of further extractions of video data and investigations of LSTM modelling results.