Inferring driving trajectories based on probabilistic model from large scale taxi GPS data

Abstract Use of taxi vehicles as mobile sensors to collect traffic information has become an important and emerging approach to relieve congestion. Global Positioning System (GPS) trajectory data allow for abundant temporal and spatial information to be collected that reflect the mobility and activity of drivers. In this paper, we present a probabilistic model to predict driving trip paths based on a Hidden Markov Model (HMM). The first step in our approach was to detect the stays or destinations by using an improved algorithm based on taxi status information about whether or not the vehicle is occupied by passengers. Next, the trips between two stays were extracted and expressed as data chains for applying the learning scheme in a predictive model. A data-driven approach based on an HMM was trained with trips from a period of three months and then the model was used to predict the future links on which the vehicle may travel. A Linear Motion Function (LMF) was then utilized to infer the taxi position on the predicted link. Furthermore, a learning algorithm was used to identify incorrect links for a given path. Finally, the effectiveness of the improved model was tested. In the testing process, samples that contained different trips with different taxi statuses (i.e., occupied and non-occupied) were considered, and the Relative Accuracy (RA) was applied as a measure of effectiveness to evaluate prediction performance. The prediction results verified that the proposed method is an accurate and feasible potential approach that can be used to estimate driving paths in future time periods.