From saliency based image features towards semantic mapping

摘要

Recent research progress enables mobile ground robots and UAVs to create maps of larger and less restricted environments. However, since these maps build upon landmarks designed for computers, like image keypoints, the benefit for human operators is limited. To close this gap, we present mapping results based on automatic extraction and matching of image features that consider the natural perceptual experience of human operators. The extraction mechanism is borrowed from the early human visual system and finds image areas (called proto-objects) that are likely to provoke bottom-up visual attention of a human observer. This work builds upon the neuroscientific Itti-Koch model. We introduce a new approximation of the normalization procedure of this model to achieve real-time performance. For proto-object matching, we propose a combined descriptor of color, texture and intermediate results of the saliency computation process. To evaluate the new landmarks, results in the areas of object-recognition and simultaneous localization and mapping are presented. Keywords Mobile Robotics, Mapping, Saliency, Proto-Objects, Visual Landmarks, SLAM