Vehicles driving behavior recognition based on transfer learning

Due to the complexity of experiments to test driving behaviors and the high cost of data collection for some types of vehicles, e.g., heavy-duty freight vehicles, it is normally hard to develop a model with a small size of samples for higher performance to correctly recognize driving behavior patterns. This paper proposes an effective recognition method based on the Convolutional Neural Network (CNN) and transfer learning. Firstly, a CNN model was constructed that was coupled with multi-source data fusion, natural driving GPS data, and drivers’ facial expression data of online car-hailing, to recognize the feature maps of five driving behavior patterns, including acceleration, deceleration, turning, lane changing, and lane keeping. Secondly, the transfer learning algorithm was employed to fine-tune the pre-trained CNN model parameters with few natural driving data samples of heavy-duty freight vehicles, where data collection is traditionally very difficult. The experiments showed that this transferred model yields a higher performance with an accuracy score of 0.80 than the non-transferred one with an accuracy of 0.64 only. Additionally, such a transferred model converged very fast with a lower training cost. After only 1,000 training epochs, its performance is much better than that of the non-transferred model after 5,000 epochs. The results demonstrated that transfer learning is an effective potential method for driving behavior recognition and other similar studies where the sample size is relatively small due to various reasons.