System for Hands-on Steering Wheel Detection Using Machine Learning
Examensarbete för masterexamen
Complex adaptive systems (MPCAS), MSc
More cars are getting semi-automatic driving capabilities and on the way towards full automation, the driver is still responsible for the car. An important aspect of controlling the car is holding the steering wheel which is supposed to be done at all times, even during semi-automatic driving. To check that the driver is controlling the vehicle and holds the steering wheel, a system can be created using inputs from the steering gear sensors or a camera. Two different solutions using these inputs are implemented and tested to see if machine learning can be used to differentiate between hands on and hands off situations. The systems are also evaluated running online on a Raspberry Pi single board computer where the efficiency of the systems is considered, as well as performance. The evaluation of the systems shows that the sensor version can be updated frequently at 5 Hz, while being robust when the driver is not intentionally tricking the system by hanging a weight on the steering wheel. The camera model is limited in update frequency by the library used for running the neural network in C++. The neural network had to be downsized in order to make the system update frequently enough to detect change of state. This meant that the system wasn’t able to detect some situations as well as the system which was evaluated on offline data. Changing libraries for the neural network would potentially solve this problem and make the online version perform more similarly to the offline system.
Hands-off detection , Machine learning , State-space model , Kalman filter , Object classification