A Method of Snowfall Detection using Spatio-Temporal Data

Abstract

This paper describes a method of snowfall detection, which is mainly utilized in road management, using spatio-temporal data. Our method extracts snowflakes taken with a flat target painted black for the background of images. First, it extracts bar regions in time differential images. Next, the extracted bar regions are classified into two categories: (1) snowflake candidates, in which the bar regions are a regular distance apart in spatio-temporal data, and (2) noise, in which the regions are near. Finally, the method judges snowfall conditions according to feature parameters such as the spatio-temporal distribution for snowflake candidates. Experimental results show that our method can detect snowfall even at low intensity and the error rate is estimated to be very low, and also indicate that this method has the potential of application to road images in which the target does not exist.