Truck platooning has great potential for reducing transport costs by lowering fuel consumption and increasing traffic efficiency. The short time headway between trucks in a platoon makes detecting the behaviour of other road participants essential for safety. Current safety controllers rely only on the traffic situation at the same instant, but accurate predictions of traffic behaviour are necessary to optimize the distance between the trucks and use the full potential of truck platooning in a safe way.

This study aims to show the potential of applying machine learning techniques to in-vehicle sensor data for predicting a cut-in manoeuvre by a passenger car. We have trained several algorithms, ranging from linear regression to Support Vector Regression and LSTM neural networks, on a dataset of naturalistic driving that contains 146 cut-ins. The results were compared to a benchmark of linear extrapolation under the assumption of a constant speed of the passenger car.

The results show that many machine learning algorithms are no viable alternative to the constant speed benchmark, with the exception of linear methods and Support Vector Regression. Further development of the Support Vector Regression algorithm in a direct-recursive hybrid forecast framework (dubbed dr-SVR) shows improvement of the error in the longitudinal distance and speed with more than 40% compared to the benchmark. Testing the trained algorithm on a truck platooning dataset shows an improvement of 15%.

The dr-SVR model has the potential to improve the safety of truck platooning by predicting the behaviour of passenger cars after a cut-in. More training data, especially including rare outliers and cut-ins representative for merges in a truck platoon, are needed to improve the accuracy and make the method suitable for application in safety controllers in the platooning trucks.