Establishing Style-Oriented Driver Models by Imitating Human Driving Behaviors

Abstract

Driver modeling is important for both automobile industry and intelligent transportation. One of its key topics has been studied in this paper, i.e., the style-oriented driver modeling for speed control, and a modeling scheme based on distal learning control and real-world vehicle test data (VTD) is proposed to make this possible. The driver model adapts to be an inverse model of the vehicle at run, which is accomplished under the distal guidance of the discrepancy between the actual and the desired vehicle speed. To tackle the divergence and local mutability of real-world vehicle data, the partly connected multilayered perceptron (PCMLP), which is a locally designed neural network, is utilized to fulfill the imitating of human operations on gas or brake pedals during real-world driving. The FTP-75 driving cycles are borrowed to test the established driver models, and driving styles of the original human drivers are retained and reproduced while accomplishing the speed following task. Simulations are conducted to verify the effectiveness of the proposed scheme.