Date of Award
Doctor of Philosophy (PhD)
Computer and Information Science
Integration of human semantics plays an increasing role in robotics tasks such as mapping, localization and detection. Increased use of semantics serves multiple purposes, including giving computers the ability to process and present data containing human meaningful concepts, allowing computers to employ human reasoning to accomplish tasks.
This dissertation presents three solutions which incorporate semantics onto visual data in order to address these problems. First, on the problem of constructing topological maps from sequence of images. The proposed solution includes a novel image similarity score which uses dynamic programming to match images using both appearance and relative positions of local features simultaneously. An MRF is constructed to model the probability of loop-closures and a locally optimal labeling is found using Loopy-BP. The recovered loop closures are then used to generate a topological map. Results are presented on four urban sequences and one indoor sequence.
The second system uses video and annotated maps to solve localization. Data association is achieved through detection of object classes, annotated in prior maps, rather than through detection of visual features. To avoid the caveats of object recognition, a new representation of query images is introduced consisting of a vector of detection scores for each object class. Using soft object detections, hypotheses about pose are refined through particle filtering. Experiments include both small office spaces, and a large open urban rail station with semantically ambiguous places. This approach showcases a representation that is both robust and can exploit the plethora of existing prior maps for GPS-denied environments while avoiding the data association problems encountered when matching point clouds or visual features.
Finally, a purely vision-based approach for constructing semantic maps given camera pose and simple object exemplar images. Object response heatmaps are combined with known pose to back-project detection information onto the world. These update the world model, integrating information over time as the camera moves. The approach avoids making hard decisions on object recognition, and aggregates evidence about objects in the world coordinate system.
These solutions simultaneously showcase the contribution of semantics in robotics and provide state of the art solutions to these fundamental problems.
Anati, Roy Corrado, "Semantic Localization and Mapping in Robot Vision" (2016). Publicly Accessible Penn Dissertations. 1587.