Deep Learning Approach for Spatial Extension of Traffic Sensor Points in Urban Road Network

Abstract

Real-time monitoring of road traffic variables is essential for any effective control strategy in Intelligent Transportation Systems. Network-wide monitoring has increased importance in the current and future panorama due to the verge of adoption of smart mobility technologies, i.e. monitoring all links in a network is a general desired goal. However, installation and maintenance of sensors across the whole network are not cost-effective. Therefore, traffic networks are frequently suffering from the lack of well-operating and reliable traffic detectors. The paper proposes the employment of neural networks based models to virtualize the measurements on road links without detectors. The proposed method applies the measurements of monitored links as input to the deep learning model in order to estimate virtual measurements on the unmonitored road links. Several neural network models differing in architecture (Artificial Neural Network, Time Lagged Neural Network and Long Short Term Memory Neural Network) have been implemented and their hyper-parameterization were optimized using Bayesian search. The prediction techniques were developed and tested by using microscopic road traffic simulation.