Classification of Stereo Images from Mobile Mapping Data Using Conditional Random Fields

Abstract

We propose a new method for the context-based classification of point clouds from stereo images using Conditional Random Fields (CRF). The classification is based on segments as nodes for the CRF. The segmentation is conducted on the image and is transferred to the 3D point cloud obtained by image matching. This allows the computation of 3D features additionally to the image features as well as the definition of realistic adjacencies between the segments in object space. We also propose a variant of the contrast-sensitive Potts model that is tailored for the contextual classification of point cloud segments. The evaluation of our method is performed on stereo sequences of a benchmark dataset, recorded in an urban area, and yields results with an overall accuracy of more than 90%. Moreover, we can show that the consideration of contextual information during the classification leads to an improvement of the overall accuracy.