Learning Algorithms for driver attitude determination

Abstract

Driver assistance systems such as Adaptive Cruise Control or Lane Keeping Assist are nowadays part of the driving experience. The research towards these technologies aims to increase the driver’s safety and comfort. These systems are rapidly becoming more functional with the help of additional sensors and technology. By default, they are tuned by test engineers according to fixed requirements, or in other words “one size fits all” model. However, to increase acceptance of these assist functions, drivers want to recognize themselves in the way the car is driven. A generic high-level framework for personalized driver-assistance systems is proposed in this study. The framework is implemented with a machine learning method known as neural networks on two different subtasks, to learn the longitudinal acceleration and headway distance preferred by the driver; and to predict the future steering wheel inputs provided by the driver. Experiment has been performed on predicting the driver’s steering wheel inputs, using a long short-term memory neural network and a multilayer perceptron. The data used in this study has been gathered through simulations with IPG CarMaker. The results show that driver’s steering wheel inputs can be predicted accurately up to 10 seconds. Furthermore, the multilayer perceptron was trained additionally by simulating real time learning. This real time learning did not prove to increase the performance significantly. Keywords: