Abstract
RGB-D object recognition for semantic mapping
Semantic mapping is the act of enhancing a standard occupancy map with high-level information. This additional knowledge can help an autonomous robot in making better-informed decisions. Knowing the nature of an obstacle might allow a robot to judge that pushing it is the best way to reach a goal position that would otherwise be unachievable. We present our approach to object detection and instance recognition with an RGB-D camera in indoor environments. The detection of objects is done by a geometric segmentation of the scene and the recognition by using supervised learning to fuse color, texture, shape and position information. The fusion helps dealing with occlusions, partial views and changes in appearance that occur in robotic experiments. This allows us to produce maps with reliable semantic information such as object identity tags, properties and snapshots.