Predicting Traffic Phases from Car Sensor Data

Abstract

This research aims to predict traffic density using driver behaviour as collected from the CAN bus. The hypothesis is that driver behavior is influenced by traffic density in such a way that an approximation of the traffic density can be determined from changes in the driver behavior. Machine learning will be employed to correlate a selection of commonly available sensors on cars to the traffic density. Challenges in the processing of the data for this purpose will be outlined. The data for this study is collected from five passenger cars and nineteen trucks driving the A28 highway in Utrecht region in the Netherlands. The results show that traffic density can be detected using driver behaviour. An overall accuracy of over 95\% is achieved with a precision of 92%. The recall rate however is low most likely caused by overfitting due to the unbalanced data set. The results still look promising and more training data should improve the results.