Soft Sensing Methods for the Generation of Plausible Traffic Data in Sensor-less Locations

The deployment of sensors in urban and interurban roads is an eminently important aspect of traffic supervision and management. However, the decision to deploy a new sensor is often subject to budgetary and operational constraints. For this reason, several alternatives to sensing are emerging aimed at reducing costs or simplifying the traffic data collection process. Among them, techniques to generate artificial data departing from a known distribution are already in use in the traffic context, allowing practitioners to augment available data, estimate future behaviors of a complete traffic network, or impute missing data. Nonetheless, the idea of generating new data from non-sensorized locations has been scarcely investigated to date, as it poses considerable challenges such as the lack of an explicit distribution to be learned, and the inherently multimodal nature of traffic data. This paper finely pulses the extent of such challenges, and proposes a selection of data generation approaches based on Generative Adversarial Networks and deep learning regression that could be used in practical traffic scenarios. To do so, we propose taking advantage of well-known spatial-temporal relations of nodes of a traffic network. With the limitations dictated by the lack of previous data, generative models are expected to obtain plausible data, with yet coarse performance metrics. Above all, performance is not the main goal of this work, but to identify and understand the main data-engineering characteristics that are critical for synthetic traffic data generation.