A Two-Stage Learning Framework for Driver Lane Change Intention Inference

Abstract

Abstract With the rapid development of intelligent vehicles, human drivers are sharing the control authority with the automation functionalities. The mutual understanding between the intelligent vehicle and the human driver is the key to the efficient multi-agent teaming and the collaborative driving system. In this study, a two-stage learning framework for spatial-temporal driver lane change intention inference is developed. The two-stage learning framework contains two separate parts. First, an operational level driving behavior recognition system based on the deep convolutional neural network (CNN) is developed to recognize the driver behaviors, including mirror checking and normal driving. Then, based on the driver behavior features from the CNN, the sequential feature vector will be used for the construction of driver intention inference model based recurrent neural networks (RNN) and long short-term memory (LSTM). The proposed model achieved a state-of-the-art result on the driver lane change intention inference task. Based on the naturalistic driving dataset, the model achieved over 91% accuracy for lane change and lane-keeping intention prediction. The two-stage learning framework can significantly increase the model flexibility and accuracy, which makes it easier to be implemented and updated in intelligent vehicles.