A new algorithm using spatial temporal domain in route inference

\noindent The research herein addressed travel path inference, using sparse global positioning system (GPS) samples, in addition to deep neural networks, which were applied for traffic flow estimation in the transportation domain. To conduct travel path inference using sparse GPS trajectory data, the GPS samples must first be localized on a road network, and then the path that a driver may have been taking must be reconstructed using a low rate of sampled GPS observations. The present work proposed a novel method to model travel path inference, in the temporal and spatial domains, as an optimization problem, in addition to a novel hybrid hidden Markov model (HMM), which made use of a uniform cost search (UCS)-like novel combinatorial algorithm. \noindent The improvements that were made herein when compared to previous studies in the literature that have used HMM based methods comprised the following: 1) For travel path inference between matched GPS positions, the hybrid HMM algorithm that was proposed in this research conducted an evaluation of all of the prospective paths, with the aim of determining the most probable path for the spatial and temporal domains. In contrast to this, the work in previously conducted studies has consisted of either interpolated trajectories or matched GPS positions that were connected via the shortest path assumption, which may have been incorrect, especially in the case of urban road networks. 2) The hybrid HMM algorithm that was proposed in this research only utilized legal speed limits to evaluate the inconsistency in the temporal domain if a situation arose in which there were insufficient historical average speed data. If this was not the case, then the historical average speed that was computed from the data was utilized. In the experiments conducted herein, which used real data sets, it was clearly seen that the proposed algorithm was able to perform more accurately than the state-of-the-art VTrack algorithm, especially in instances in which the GPS data were infrequently sampled. \noindent For the application of deep neural networks in traffic flow estimation, the problem of hot road detection, which involves the estimation of the road density at a future time based on previous observations, was selected. Sensors on the road networks record number of cars passing through at certain intervals. By modelling these recordings as a sequence of observations in the temporal domain, variants of deep neural networks that are designed for sequenced input were applied such as long short time memory (LSTM), gated recurrent units (GRU) as well as typical stacked auto encoders (SAEs). The results were analyzed and the performance of the different variants of the neural networks were compared. The experiments in the traffic flow estimation showed that when the LSTM, GRUs, and SAEs were trained on a single sensor for durations of 30, 60, 90, and 120 min of training data, the models were powerful enough to estimate the next 5 min of traffic state.