Model-based detection and segmentation of vehicles for intelligent transportation system

In this paper we present a method to detect vehicles on the road for intelligent transport application. In the proposed method, statistical property of background pixels are defined, where foreground objects of a current frame are detected by subtracting the current frame w.r.t. background color model and background edge model. Color and edge are considered here as these two features provide complementary advantages. On top of this pixel-based step of subtracting pixel w.r.t. background models, region property is incorporated, where segmentation of a foreground is also subject to its appearance in spatial domain. Clustering is performed to capture local areas with high entropy indicating the presence of a foreground. A region-based speed function is proposed to evolve an active contour for a final segmentation based on region property within the contour. We observe that our strategy of integrating pixel-based and region-based modeling yields robust performance in handling challenging scenarios. The experiment of this work reports results of the proposed method when evaluating it on surveillance videos recorded by a live system.