Detecting Road Users at Intersections Through Changing Weather Using RGB-Thermal Video

Abstract

Abstract. This paper compares the performance of a watch-dog system that detects road user actions in urban intersections to a KLTbased tracking system used in traffic surveillance. The two approaches are evaluated on 16 hours of video data captured by RGB and thermal cameras under challenging light and weather conditions. On this dataset, the detection performance of right turning vehicles, left turning vehicles, and straight going cyclists are evaluated. Results from both systems show good performance when detecting turning vehicles with a precision of 0.90 and above depending on environmental conditions. The detection performance of cyclists shows that further work on both systems is needed in order to obtain acceptable recall rates.