Segmentation of urban traffic scene based on 3D structure

This paper proposes an approach for image segmentation of the urban traffic scene captured from a car-mounted camera. First of all, an improved SIFT feature matching algorithm is adopted for extracting 2D keypoints of the scene. Tracking the 2D keypoints generates the 3D points clouds that can estimate the 3D world structure and motion features. And then Multiple Kernels Support Vector Machines (MKSVM) is employed for sematic segmentation based on motion-derived 3D structure and SIFT features. Experiments show the efficiency and the relevancy of our approach.