The purpose of this research was to investigate object classification algorithms for the application of wheelchair interaction with the environment for the cognitively impaired wheelchair user. Towards this end, top performing object classification algorithms were trained on images of the target object classes (chair, dresser, and sink/washbasin) obtained from the internet and tested on images of the target object classes obtain in the home and patient room environments; these algorithms were Locality-constrained Linear Coding (LLC) [1], Kernel Descriptors (KDES) [2], and Hierarchical Matching Pursuit (HMP) [3]. It was found that HMP achieved the highest over classification accuracy (71.3%) in the home environment and LLC achieved the greatest accuracy (85.0%) in the patient room environment. This research also sought to investigate the potential of active learning to improve upon the obtained classification performance. A maximum mean classification accuracy of 98.6% was achieved when active learning was applied.
| Identifer | oai:union.ndltd.org:TORONTO/oai:tspace.library.utoronto.ca:1807/43280 |
| Date | 09 December 2013 |
| Creators | Oramasionwu, Paul |
| Contributors | Mihailidis, Alex |
| Source Sets | University of Toronto |
| Language | en_ca |
| Detected Language | English |
| Type | Thesis |
Page generated in 0.0019 seconds