Driving behaviour characterisation by using phase‐space reconstruction and pre‐trained convolutional neural network

Driving behaviour analysis is important for both intelligent transportation and public security. The authors propose to characterise driving behaviours by using the phase-space reconstruction (PSR) and the pre-trained convolutional neural network (CNN). PSR is first applied to the raw vehicle test data (VTD) to obtain the reconstructed trajectories. Second, the corresponding feature vectors are acquired by using the pre-trained CNN. Third, the t -distributed stochastic neighbour embedding (t -SNE) algorithm is applied to the feature vectors to validate their characterising ability. Finally, an index is proposed based on the aforementioned feature vectors for quantitative evaluation, i.e. driving style recognition and abnormal driving detection. Simulations are conducted to verify the effectiveness of the proposed scheme.