Context Based Object Recognition for Mobile Platforms, 10-R8065

Printer Friendly Version

Principal Investigator
Shaun Edwards

Inclusive Dates:  04/27/09 – 08/27/09

Background - SwRI explored the benefits of integrating perception methods developed under a previous internal research project, "Ontologies for Object Recognition," for use on small mobile robots. Indoor surveillance and reconnaissance was the application focus of this research. For these applications, which include highly dynamic and complex indoor environments, highly accurate object recognition is required. Current approaches for small mobile robots in indoor environments have focused on simple navigation with limited object recognition capabilities. Extending this research to include accurate and diverse object recognition would allow for better characterization of the indoor environment.

Approach - The goal of this project was porting or adapting context-based object recognition to a mobile robot architecture. The tasks included porting of algorithms to C++ so that they could be easily included with the robot software, replacement of the ontology used in the original research with the ConceptNet knowledgebase, and image processing and segmentation algorithms for stereo image data. Because of the many tasks required, standard open source tools were leveraged for this effort to make quick progress and identify areas that will require improvement. The resulting recognition algorithms were tested in an indoor office and were capable of recognizing twenty-seven different object types commonly found within an indoor environment. The classification rates of classical object recognition methods and the SwRI context-based methods were compared to determine the relative effectiveness of context for indoor environments.

Accomplishments - The SwRI context-based object recognition methods were successfully ported to a mobile robot architecture and tested in an indoor environment. The results of this testing found that context-based methods did not improve object recognition when compared to other methods because of a lack of contextual knowledge in the ConceptNet knowledge base. Follow-on research is being proposed to a U.S. Department of Defense client to enhance the ConceptNet knowledge base and to address other research areas identified as part of this program.

2009 Program Home