Visualizing Road Appearance Properties in Driving Video

Abstract

With the increasing videos taken from driving recorders on thousands of cars, it is a challenging task to retrieve these videos and search for important information. The goal of this work is to mine certain critical road properties in a large scale driving video data set for traffic accident analysis, sensing algorithm development, and testing benchmark. Our aim is to condense video data to compact road profiles, which contain visual features of the road environment. By visualizing road edge and lane marks in the feature space with the reduced dimension, we will further explore the road edge models influenced by road and off-road materials, weather, lighting condition, etc.