Recurrent End-to-End Neural Network Design with Temporal Dependencies for Model-Free Lane Keeping Systems

Abstract

Recently, many kinds of research on advanced driver assistant system (ADAS) for driver convenience and active safety has been actively carried out. In the conventional lane keeping system (LKS), the control input of closed-loop feedback system is calculated based on the lateral motion model. The biggest problem of the model-based approach is that there are various parameters in the model dynamics and the unknown values are computed through various assumptions. In this paper, we propose a model-free lane keeping system based on long short term memory (LSTM) which considers the time sequential information of data. The dataset was collected by various sensors and reshaped for the training. We can obtain the steering command directly by the end-to-end recurrent network. To validation, the LKS simulation was conducted and we observed that the proposed method is more accurate than the lane keeping by feedback control.