- Title
- Automatic landmark detection for localisation and navigation
- Creator
- Bhatia, Shashank
- Relation
- University of Newcastle Research Higher Degree Thesis
- Resource Type
- thesis
- Date
- 2015
- Description
- Research Doctorate - Doctor of Philosophy (PhD)
- Description
- The objective of this research work is to advance the process of integrating mobile service robots in public environments. With the growth in use of service robots in the complex world of humans, they are required to sustain long-term autonomy while possessing natural scene understanding capability. Robotics researchers have devised robust mechanisms to impart the capacity for localisation/way finding and navigation to these robots. However, the smooth integration of such service robots into existing industry, healthcare and warehousing environments is yet to be seen. Often, the addition of a robot worker requires additional infrastructure to support the robot’s way-finding process. As a result, these robots are restricted to certain fixed parts of the building where the infrastructure has been installed. An example of such a service robot is the Swisslog Telelift commissioned at the Royal North Shore Hospital, Sydney. The service robot is an Automated Ground Vehicle that has been commissioned to transport food, linen and supplies across various departments of the hospital. This thesis contributes towards enabling a smoother integration of such service robots, and reconditioning their way-finding capacity to support long-term autonomy. The enhancement is specifically focused towards minimising the requirements for additional markers and beacons that act as artificial landmarks to help in the task of localisation. This is achieved by developing methods of automatically identifying salient locations - landmarks - based on their 3D structure, presence of salient objects, and by estimating a measure of surprise encountered on experiencing the location. A methodology for evaluating a 3D model of a building for the presence of landmarks is proposed. 3D isovists are employed to perceive the structure of the environment, which is then processed further to extract the gist of the entire structure. The statistical summaries extracted from these 3D isovists are compared and analysed to identify those that are different from others within a small neighbourhood (local saliency), and throughout the entire building (global saliency). The developed method is converted into an empirical methodology involving extraction of the 3D isovists from a simulated walk-through experience of a 3D model of the corresponding building. Hypothetical architectural models of two canonical house designs: the Villa Savoye at Poissy (near Paris, France) by Le Corbusier and the Dana-Thomas House at Springfield (Illinois, USA) by Frank LloydWright are employed to demonstrate the use of the proposed technique. The methodology serves two purposes: as a tool for evaluation of plans by architects for the presence of structurally variant locations, and as a method to evaluate the structural saliency of a location as detected by a robot. When the structure of the environment is monotonous, humans tend to remember a location by associating it with the presence of any salient object. Aiming to imitate this capacity, a top-down saliency evaluation method is proposed, which enables a robot to isolate objects from the background built environment (often included in a sensory-percept such as lidar). Residual object data from the percept is then analysed to identify the most salient object present, based on the curvature and silhouette of the object. The proposed technique is demonstrated to have both repeatability and robustness to viewpoint changes. Multiple sensors are often employed by robots to perceive a holistic view of the environment. Each sensor contributes a different type of information. In such scenarios, evaluating saliency becomes a complex problem. Instead, a memory-based saliency mechanism is proposed. The design of a heteroassociative computational memory and its extension in the form of a multidirectional associative memory model are presented in this thesis. The memories are equipped with the capacity to evaluate input patterns for the amount of artificial surprise they stimulate. The philosophy used for learning and training is as follows: a novel input that causes more surprise is highly valuable. The memory models start from no prior knowledge, perceive multiple input patterns, and selectively store those patterns that do not match the immediate prior belief. The memory models are designed to support domain-independence and probabilistic techniques are employed to establish differences between prior and posterior belief, accounted as surprise. A framework suitable for integrating the proposed memory model in the robotics domain is proposed. The progress made in this research supports architects to plan the design of public buildings and enhancing the support that these buildings offer for way finding by both humans and robots. Additionally, the techniques also find utility in a robot localisation scenario for identifying candidate perceptual landmarks.
- Subject
- landmarks; localisation; navigation; saliency
- Identifier
- http://hdl.handle.net/1959.13/1063105
- Identifier
- uon:17203
- Rights
- Copyright 2015 Shashank Bhatia
- Language
- eng
- Full Text
- Hits: 1143
- Visitors: 1397
- Downloads: 357
Thumbnail | File | Description | Size | Format | |||
---|---|---|---|---|---|---|---|
View Details Download | ATTACHMENT01 | Abstract | 322 KB | Adobe Acrobat PDF | View Details Download | ||
View Details Download | ATTACHMENT02 | Thesis | 17 MB | Adobe Acrobat PDF | View Details Download |