Deep Neural Network Gaussian Process Regression Method for End-to-end Driving Behavior Learning

Abstract

For dynamic complex driving tasks of autonomous vehicle, traditional driving behavior learning methods still need to be improved in terms of learning accuracy and generalization performance. A convolutional neural network gaussian process regression (CNN-GPR) method for driving behavior learning is proposed to tackle these problems that full connection layers of end-to-end convolutional neural networks (CNN) have limited generalization ability and easily converge to local optimization. At the same time, the PilotNet-GPR algorithm based on the CNN-GPR method is designed. The gaussian process regression (GPR) method with global mapping capability and better generalization performance is used to improve fully connected layers of the end-to-end CNN in order to complete the mapping from features extracted to driving actions more efficiently for the proposed CNN-GPR method. In addition, the long short-term memory network (LSTM) is added into CNN-GPR method, and a convolutional long short-term memory network gaussian process regression (CNN-LSTM-GPR) method of driving behavior learning with time-sequential images is proposed in order to further improve the accuracy of driving behavior learning through time-sequential information. This method utilises the gaussian process regression method to improve structures of fully connected layers in the cascaded deep neural network (CNN-LSTM) in order to make a more efficient learning approximation between features extracted by the CNN-LSTM and driving actions. Verification experiments on the Apollo end-to-end dataset for autonomous driving show that the proposed CNN-GPR method can further improve the imitation accuracy for end-to-end driving behavior learning and promote generalization performance of learned models compared with the PilotNet method. compared with related end-to-end driving behavior learning methods based on single image under the same condition, the proposed CNN-LSTM-GPR method can make full use of time-sequential information of images, which leads to smaller imitation errors. Additionally, it can further enhance the learning accuracy and demonstrates a more satisfying generalization performance compared with the CNN-LSTM method.