Semantic Segmentation of Urban Environments into Object and Background Categories

Advancements in robotic navigation, object search and exploration rest to a large extent on robust, efficient and more advanced semantic understanding of the surrounding environment. Since the choice of most relevant semantic information depends on the task, it is desirable to develop approaches which can be adopted for different tasks at hand and which separate the aspects related to surroundings from object entities. In the proposed work we present an efficient approach for detecting generic objects in urban environments from videos acquired by a moving vehicle by means of semantic segmentation. Compared to traditional approaches for semantic labeling, we strive to detect variety of objects, while avoiding the need for large amounts of training data required for recognizing individual object categories and visual variability within and across the categories. In the proposed approach we exploit the features providing evidence about widely available non-object categories (such as sky, road, buildings) and use informative features which are indicative of the presence of object boundaries to gather the evidence about objects. We formulate the object/non-object semantic segmentation problem in the Conditional Random Field Framework, where the structure of the graph is induced by the minimum spanning tree computed over 3D reconstruction, yielding an efficient algorithm for an exact inference. We carry out extensive experiments on videos of urban environments acquired by a moving vehicle and compare our approach to existing alternatives.