Acquiring Semantically Meaningful Models for Robotic Localization, Mapping and Target Recognition.

摘要

The goal of this proposal is to develop novel representations and techniques for localization, mapping and target recognition from videos of indoors and urban outdoors environments. The proposed techniques will facilitate enhanced navigation capabilities by means of visual sensing and enable scalable, long-term navigation and target detection in outdoors and indoors environments. The attained representations will also be applicable towards human-robot interaction, enhancement of human navigational and decision making capabilities and provide compact semantically meaningful summaries of the acquired sensory experience. The proposed representations will be governed by principles of compositionality, facilitate bottom-up learning, enable efficient inference and could be adapted to a task at hand. The main novelty of the approach will be the use both 3D and 2D geometric and photometric cues computed either from video sequence or from novel RBG-D cameras, which provide synchronized video and range data at frame rate. Video poses challenges related to more extreme variations in viewpoint and scale, dramatic changes in lighting and large amount of clutter and occlusions, but also enables computation of 3D structure and motion cues, which can aid segmentation and recognition of object and non-object categories. As a part of this proposal we have developed techniques for semantic labeling of outdoors and indoors environments using photometric and geometric cues from video. The proposed approach is informed by novel features and representations for learning models of objects and non-object categories from video, works effectively with multiple sensing modalities and can be deployed on static frames as well as video in a recursive setting. We have tested the approach extensively on benchmark sequences of indoors and outdoors environments.