Automated driving is gaining more and more interest in the recent years. In order to drive safely, automated vehicles reconstruct the environment using mainly the information coming from the in-car sensors. By using this reconstruction a prediction of the future states of the surrounding vehicles can be computed, which in turn is used to decide which manoeuvre and accompanying path to accomplish. This problem is known as motion prediction. Whilst physics-based models perform well on a short horizon, machine learning has the potential to predict a more accurate motion on a longer horizon, especially if the manoeuvre of the other road user is known in advance (manoeuvre-based prediction).
In this paper a hybrid approach is proposed that consists of an intention of manoeuvre predictor, a physics-based motion predictor and a manoeuvre-based motion predictor.
The vehicles around the host vehicle are continuously tracked. The intention of manoeuvre predictor, based on Support vector machines (SVM), computes the probability for each surrounding vehicle of changing lane or of staying in lane. In addition, a kinematical model which assumes a constant turn rate and velocity (CTRV) is used to predict the trajectories. Once the intention of manoeuvre is known, the manoeuvre-based motion model, based on machine learning algorithms as Gaussian Processes and SVR, predicts the lane change or lane following trajectories.
The models are trained using a collection of cut-ins manoeuvres from 60 hours of naturalistic driving. In the end, the physics-based and the manoeuvre-based motion predictions are merged together by a weighting function. The models were validated with cross-validation and the performance and the integration between sub-modules was tested in a Hardware In the Loop (HIL) environment. The models are capable of detecting the intention of a surrounding vehicle of changing lane with a positive predictive value of 82% 1.2 second before it crosses the lane marker. The combination of SVR and CTRV is capable of predicting well for shorter and longer horizons, keeping the advantages of both methods. The combined model predicts the longitudinal distance and the lateral distance with an error that is 50% lower than the one using the physics-based model, after 4s and an even better performance on shorter horizons in comparison with SVR.
The presented approach is capable of predicting the motion of the other road users in a standard situation. In order to handle more sophisticated scenarios, the road information should be used for training. The training set needs to be extended for better results and the models need to be validated on safety-critical scenarios.
A hybrid approach for predicting the motion of vehicles from a host vehicle perspective is presented in this work. A combination of machine learning and physics-based models is used to enhance the accuracy of the prediction in shorter and longer horizons. The information coming from the prediction module can be used path planning of (partly) automated vehicles.
The results and the integration in the HIL environment show great potential to allow autonomous driving to go to higher levels of automation.