Research

2025

Beverley Gorry, Dr. Tobias Fischer, Prof. Michael Milford, Dr. Alejandro Fontan
Available on arXiv.

Abstract:
Effective monitoring of underwater ecosystems is crucial for tracking environmental changes, guiding conservation efforts, and ensuring long-term ecosystem health. However, automating underwater ecosystem management with robotic platforms remains challenging due to the complexities of underwater imagery, which pose significant difficulties for traditional visual localization methods. We propose an integrated pipeline that combines Visual Place Recognition (VPR), feature matching, and image segmentation on video-derived images. This method enables robust identification of revisited areas, estimation of rigid transformations, and downstream analysis of ecosystem changes. Furthermore, we introduce the SQUIDLE+ VPR Benchmark-the first large-scale underwater VPR benchmark designed to leverage an extensive collection of unstructured data from multiple robotic platforms, spanning time intervals from days to years. The dataset encompasses diverse trajectories, arbitrary overlap and diverse seafloor types captured under varying environmental conditions, including differences in depth, lighting, and turbidity.

2024

Beverley Gorry, Dr. Tobias Fischer, Dr. Alejandro Fontan
In this project, I applied state-of-the-art techniques in robotic vision to high-speed environments, collaborating with the QUT Motorsport (QUTMS) team to assess photometric and feature-based autonomous navigation methods on the racetrack. This project not only deepened my technical expertise but also taught me invaluable lessons in collaboration and leadership. Working closely with the QUTMS team required me to translate complex concepts into actionable strategies that supported their goals. By effectively balancing technical research and practical implementation, I developed a robust skill set in research analysis, critical thinking, and teamwork.

2023

Beverley Gorry, Dr. Tobias Fischer
Summer research project undertaken as part of the QUT Vacation Experience Research Scheme (VRES).

Abstract:
Visual place recognition (VPR) refers to a system’s ability to identify specific locations using visual information. With applications in fields including autonomous vehicles, environmental monitoring, and search and rescue operations, being able to reliably determine a robot’s position in its environment is essential for VPR. Event cameras offer low power consumption and low bandwidth requirements, coupled with high dynamic range and minimal motion blur. These attributes render them advantageous for the VPR task, where changes in speed and textured environments can compromise the consistency of event detection outcomes.
In current research, event-based place recognition has been explored by fixing the time interval for event collection and fixing the number of events recorded in each segment. While these strategies allow the system to strike a balance between capturing sufficient temporal information and maintaining system efficiency, there is an inherent challenge with creating a consistent representation of the environment. Fixing the time period is not robust against changes in speed, as a mobile robotic platform travelling at low speeds is unable to identify as many places in the same time period as one travelling quickly. Alternatively, fixing the number of events collected in each segment enhances performance against sparse data collection. However, system consistency is challenged in highly dynamic environments or those with varying texture.
We hypothesize that fixing the distance travelled and accumulating events using the movement of the platform in the environment will introduce robustness against variations in speed and changes in dynamics of the environment. Our aim is to explore the viability of using odometry information from an IMU for visual place recognition with varying speeds and textures in the environment. This solution is evaluated on an indoor QCR-Event-VPR dataset which was captured with a DAVIS346 camera mounted on a mobile robot. Our early results show that position estimates can be used to determine segments for event collection with a consistent number of places in each segment. Thus, gathering event information using fixed distance estimates underscores the importance of consistency in event-based place recognition.

2022

Beverley Gorry, Prof. Paul Roe
Summer research project undertaken as part of the QUT Vacation Experience Research Scheme (VRES).

Abstract:
Recording and analysing environmental sounds in an urban setting provides a means of understanding the impact of urbanisation on biodiversity. Urban ecoacoustic monitoring combined with machine learning techniques facilitates the detection of particular vocal species in an environment. This information then ensures that the distribution of biodiversity can be effectively monitored and therefore maintained. This research project aims to explore and develop a method for detecting and classifying bird species in ecoacoustic recordings.
Techniques from an investigation around detecting speech in audio recordings have been adopted to create a transfer learning-based network. This includes use of the YAMNet Audio Event classification model and Python programming to develop a model which determines whether an audio sample contains bird audio. The detection model was developed and trained using a large dataset obtained from previous research and tested on real urban ecoacoustic data. Results can also be visualised by plotting the waveform and spectrogram of processed audio. The transfer learning-based network is being modified to operate using Tensorflow Lite, so that the model can be deployed in an urban environment using a Raspberry Pi.
This project may be deployed at the Samford Ecological Research Facility to accurately quantify the density of bird species in the area. As a result, the model can be used to inform more effective methods of conserving biodiversity through understanding the current distribution.