Driver route and destination prediction

Abstract

A method is proposed for estimating driver's intended route and destination. Probabilistic Bayesian models are employed to analyze the history of driving for individuals, where data attributes are GPS traces captured during trips from fleet of cars. The proposed probabilistic model is built up in the road graph level which is associated with its corresponding destination/origin and additional data describing characteristics of each trip. The proposed prediction model is built upon destination clustering [1]. To avoid overfitting of the predictive model for multiple destinations corresponding to the same physical location, we use a modified DBSCAN method to cluster the destinations. Low computational complexity, flexibility, and simplicity of the proposed algorithms that can be adapted and trained with time series data are the main advantages of our predictive model. Preliminary results evaluated for the destination prediction and short range path prediction indicate the accuracy and reliability of the proposed method.