The framework is proposed to recognize driving intentions and to predict driving behaviors of lane changing on the highway, by using externally sensable traffic data from the host-vehicle.
The framework consists of a driving characteristic estimator and a driving behavior predictor. In this work, a driver's implicit driving characteristic information is determined through two driver behavioral models, the Intelligent Driver Model (IDM) and the Minimizing Overall Braking Induced by Lane change (MOBIL) model. The parameters of the two models are identified as the metrics of a driver's driving characteristics and are used together to recognize a driver's intention on lane change. In order to extract the implicit driver characteristic parameters from the externally observed vehicle motion data, a parameter online estimator is designed by formulating into a model fitting problem, solved by using Genetic Algorithm (GA) assisted with evolving Takagi-Sugeno (eTS) clustering method to accelerate the convergence speed and convergence quality.
For intention identification and driving behavior prediction, a five-layer Neural Network with a Long Short-Term Memory (LSTM) layer is implemented and trained, so as to absorb the online estimated driver parameters and return the future probability of each defined driving behaviors. The developed driving behavior predictor is validated by testing with the real naturalistic traffic data from Next Generation Simulation (NGSIM), which demonstrates the effectiveness in identifying the driving characteristics and transforming into accurate behavior prediction in real-world traffic situations.