Predicting Visibility Conditions at the CCTV installation site using Long Short-Term Memory

Predicting visibility conditions at closed-circuit television (CCTV) installation sites is important for enhancing driver safety in winter. The present study proposes a visibility prediction method based on long short-term memory (LSTM), which is a type of deep learning architecture. The method was trained using observed weighted intensity of power spectrum (WIPS) values as visibility information, as well as meteorological data containing precipitation intensity by the weather radar at the CCTV site. eXtended RAdar Information Network (XRAIN) in Japan observed precipitation intensity every 1 min in a 250 m square grid mesh. The present study conducts an experiment to predict WIPS values using images actually captured by the CCTV camera. We verified multiple methods with different predictors and explanatory variables, and confirmed the validity of using LSTM as the predictor and meteorological data such as WIPS values, precipitation intensity, wind speed, and temperature as explanatory variables.