Use this URL to cite or link to this record in EThOS:
Title: An investigation into full body gender recognition in images and video
Author: Collins, Matthew
Awarding Body: Queen's University Belfast
Current Institution: Queen's University Belfast
Date of Award: 2015
Availability of Full Text:
Full text unavailable from EThOS.
Please contact the current institution’s library for further details.
Gender classification at an intermediate distance is a very important and challenging topic in video surveillance. Though quite a substantial body of work exists on gender classification from faces and from gait with clean background, little work has been done on gender profiling on full human body from static images with complex background. To attack this problem, it is felt that one of the key issues is the building of robust feature representations. This thesis presents compelling feature representations which extract cues such as body shape and appearance from static pedestrian images and achieve state of the art accuracies on the limited data available. Analysis of the data confirms intuitive characteristics that a human observer would look for when classifying gender, are captured by these feature representations. Across all features examined it is clear that strong spatial constraints provide important contextual information. Equally it is shown that weighted feature combination tends to boost performance over any individual feature alone. However, a comparison of machine learning algorithms used to perform these combinations indicates that the choice of algorithm takes a back seat to the features themselves. In the case of dynamic video, a refinement of a feature type that has shown success in behaviour recognition applications is presented. Parallels are drawn between the interest points detected and point light walkers which have been shown to be sufficient for an observer to recognise the gender of a walking subject. Again, strong classification accuracy and class separation are achieved. The work also addresses the lack in the research community of available gender balanced datasets in both the static and video domains. In particular a new synchronised multi-angle video dataset of 101 subjects walking on a treadmill is presented.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available