View the presentation slides  (1.9 Mb)

Abstract

Coastal and estuarine ecosystems are under increasing pressure from multiple uses. Regular monitoring and classification of benthic habitats would greatly assist the sustainable management of these ecosystems. Traditional manual classification of benthic habitat using video and still images as source data is a time-consuming, expensive and error-prone task.

A large multi-disciplinary project has been established by the CSIRO to study multiple-use coastal environments. As part of this larger study the Tasmanian ICT Centre is working on autonomous habitat mapping.

The system we have developed operates on-board a small autonomous submersible (Starbug), and enables real-time on-board classification of images, with potential to identify individual species. The algorithms we have designed are independent of the classification task and the imaging equipment. For example, the systems may be used to detect objects rather than habitat types, or use hyper-spectral rather than visible light cameras.

The accuracy of Starbug's localisation and image quality are the main factors that determine the quality of the habitat maps constructed by our system. Image quality in this environment is typically affected by water turbidity, depth at which images are taken, light attenuation, camera focus and blur caused by motion of the submersible.

To investigate the sensitivity of our system to these factors we are developing a test-bench for our algorithms. This test-bench implements a software-based model of Starbug and refines the classification algorithm using the Gazebo multi-robot simulator.

The Gazebo software simulates a marine environment and feeds simulated sensor data (images, depth, temperature, etc.) to the actual Starbug operational software suite in real time. The combination of simulated sensor readings and actual operational software greatly enhances our ability to rapidly analyse and refine the classification algorithms.

By varying parameters related to the classification algorithm (in particular, positional probabilitic distribution and environmental quality) and observing the impact of these adjustments on the accuracy of the constructed benthic map, we can establish an understanding of the operational envelope for the system, and improve the classification accuracy.

About the speaker

Greg Timms received the BSc (Hons) and PhD degrees in physics from the University of Sydney, Australia, in 1993 and 1997 respectively. In 1997, he joined the Australian Nuclear Science and Technology Organisation where he spent five years investigating the environmental impacts of mining, focusing on the physical transport of reactants and pollutants within mine wastes. Since 2002, Greg has been with the Commonwealth Scientific and Industrial Research Organisation (CSIRO), initially engaged in research on microwave communication networks and then leading a team which developed a novel 190 GHz millimetre-wave imager in 2006. For the past year Greg has been based at the Tasmanian ICT Centre at CSIRO’s Hobart site, where he has led a team deploying a marine sensor network in the waterways of southern Tasmania; this network is beginning to provide data that will be integrated with models to assist authorities in managing the multiple uses of these waterways. Greg particularly enjoys undertaking applied research with clear benefits to industry or the community.