Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.617591
Title: Egocentric activity recognition on the move
Author: Sundaram , Sudeep
Awarding Body: University of Bristol
Current Institution: University of Bristol
Date of Award: 2012
Availability of Full Text:
Access through EThOS:
Abstract:
Advances in the design and efficiency of cameras, combined with a significant increase in computational capabilities of machines, have directly resulted in the rapid evolution of computer vision systems. Cameras, which capture all texture, objects and motion in their field of view, provide tremendous potential as wearable sensors that provide aid in daily living. Given the advantages associated with visual sensors, they have been surprisingly under-used as wearables on the move. This thesis addresses the problem of visual recognition of human activity on the move from an egocentric point of view. The problem of perceived motion in the background brought about by the use of moving cameras, is handled by a novel method for background subtraction. Once foreground motion has been disambiguated, a unique representation of actions as collections of smaller space-time volumes is presented, using which two novel methods are proposed for the recognition of egocentric and external actions. The element of context plays a. vital role in complete understanding of human activities. This aspect is addressed through the recognition of user location using a Simultaneous Localisation and Mapping system. Activities and locations are modelled in a complementary manner, such that knowledge of what happens where enhances mapping of large environments and also increases accuracy of activity recognition. The combination of action recognition and location recognition is applied for sequential recognition of activities in continuous video, which in turn opens doors to applications such as life logging and activity retrieval. Results arc shown on datasets captured using a shoulder-worn camera as a solitary sensor. Although the proposed methods outperform the state of the art in egocentric visual activity recognition, it remains a. significant first step towards truly autonomous wearable assistance.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.617591  DOI: Not available
Share: