Spatial Prior for Nonparametric Road Scene Parsing

Parsing road scene images taken from vehicle mounted cameras provides important information for high level tasks in automated on-road vehicles. In this paper we adopt the nonparametric framework for this problem and present a simple yet effective strategy to integrate spatial prior into the framework. Unlike natural scene images, road scene images in our problem typically have very stable scene layout, which motivates us to explore such layout for improving scene labeling. In particular, the spatial distribution of each semantic category is obtained from a set of previously observed data. Then, such distributions, in the form of histograms, are integrated into the nonparametric labeling framework to guide scene parsing. Compared with previous approaches, our solution is very efficient in both computation and memory usage, since there is no complicated semantic training involved. For evaluation, we collected three video datasets on three different trips and ran the proposed algorithm on all of them, both within each trip or cross trip. The experimental results show advantages of our algorithm.