Probabilistic model for vehicle trajectories reconstruction using sparse mobile sensor data on freeways

Real-world vehicle trajectory data are increasingly used in many applications such as driver behavior investigation, travel time estimation, and vehicle energy/emission evaluation. Due to the omnipresence of mobile sensors such as smartphones, real-world vehicle trajectory data can be collected at a large scale. However, most large-scaled mobile sensor data in practice are sparse in terms of sampling rate due to implementation costs. Before these data can be used in some applications, it is necessary to first reconstruct the data into high-resolution vehicle trajectories (e.g., second-by-second). In this paper, a probabilistic model for reconstructing vehicle trajectories from sparse mobile sensor data collected on freeways is presented. Prior probability distributions are used to quantify the statistics of different driving modes, e.g., acceleration, deceleration, cruising, and idling. The maximum likelihood estimation (MLE) technique is then used to reconstruct the second-by-second vehicle trajectory between consecutive sampling times. The proposed model is calibrated and validated using the NGSIM's US-101 dataset, and shows great promise in enhancing the applicability of sparse mobile sensor data. The results show that the Mean Absolute Error (MAE) of the estimation on second-by-second location and speed is 1.39 m and 0.47 m/s, respectively.