Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.713135
Title: Scene analysis and risk estimation for domestic robots, security and smart homes
Author: Dupre, Rob
Awarding Body: Kingston University
Current Institution: Kingston University
Date of Award: 2017
Availability of Full Text:
Access through EThOS:
Access through Institution:
Abstract:
The evaluation of risk within a scene is a new and emerging area of research. With the advent of smart enabled homes and the continued development and implementation of domestic robotics, the platform for automated risk assessment within the home is now a possibility. The aim of this thesis is to explore a subsection of the problems facing the detection and quantification of risk in a domestic setting. A Risk Estimation framework is introduced which provides a flexible and context aware platform from which measurable elements of risk can be combined to create a final risk score for a scene. To populate this framework, three elements of measurable risk are proposed and evaluated: Firstly, scene stability, assessing the location and stability of objects within an environment through the use of physics simulation techniques. Secondly, hazard feature analysis using two specifically designed novel feature descriptors (3D Voxel HOG and the Physics Behaviour Feature) which determine if the objects within a scene have dangerous or risky properties such as blades or points. Finally, environment interaction, which uses human behaviour simulation to predict human reactions to detected risks and highlight areas of a scene most likely to be visited. Additionally methodologies are introduced to support these concepts including: a simulation prediction framework which reduces the computational cost of physics simulation, a Robust Filter and Complex Adaboost which aim to improve the robustness and training times required for hazard feature classification models. The Human and Group Behaviour Evaluation framework is introduced to provide a platform from which simulation algorithms can be evaluated without the need for extensive ground truth data. Finally the 3D Risk Scenes (3DRS) dataset is introduced, creating a risk specific dataset for the evaluation of future domestic risk analysis methodologies.
Supervisor: Not available Sponsor: Kingston University
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.713135  DOI: Not available
Keywords: Computer science and informatics
Share: