Deep learning method for driver identification using vehicle sensor data

Abstract

With greater emphasis on user privacy protection, sensor data without biometric identifier reflects driver style and is commonly used for non-invasive identification of driver behavior. In this project, we use the sampled Nervtech dataset and choose Driver2vec and Gradient boosting decision tree (GBDT) for driver identification. The model utilizes the dominance of temporal convolutional network, wavelet transform, margin hard triplet loss, and GBDT classifiers. The idea of the model is to map the temporal inputs into a 62-dimensional embedding space, and use the trained driver embeddings as inputs to the classifier to do prediction. Driver2vec output embeddings are demonstrated to form well-differentiated clusters in T-SNE visualization. The model achieves 71.4% average pairwise accuracy on the test sets, which is higher than previous studies. After integrating lightGBM, xgboost, random forest, the model performance improves about 4% to 74.6% and the generalization ability is enhanced. This Driver2vec model can be applied in different driving environments and also ensure high accuracy prediction. After introducing 10% of “noisy” data, the pre diction accuracy of the model remains above 60%, which indicates good robustness.