Auto-Encoder LSTM for learning dependency of traffic flow by sequencing spatial-temporal traffic flow rate: A speed up technique for routing vehicles between origin and destination

Abstract

Urban transport system is a time varying network. The variation in travel time and delay in travel faced by commuters is the adverse effect of traffic congestion. Traffic information in preceding time instances contributes in analyzing traffic in succeeding instances and spatial information of traffic is required for traffic flow assessment on highways. Sequencing spatial and temporal traffic information in preceding time instance helps in estimating traffic flow in sequence in successive time instances by formalizing Sequence Convolution based auto-encoder Long Short term Memory (SCAE-LSTM) network. The objective of this work is to estimate traffic flow on highways for different origin-destination (OD) pair based on spatial-temporal traffic sequences. Hence, Spatial-TemporAl Reconnect (STAR) algorithm is proposed. The performance of STAR is investigated by conducting extensive experimentation on real traffic network of Chennai Metropolitan City. The computational complexity of the algorithm is empirically analyzed. The proposed STAR algorithm is found to estimate traffic flow during peak hour traffic with reduced complexity in computation compared to other baseline methods in short term traffic flow predictions like LSTM, ConvLSTM and GRNN. Finally, conclusions on results are presented with directions for future research.