Wildlife surveys are a key tool used to manage threatened and endangered species. Historically, these have been performed manually, however a growing number of surveys are using drones and automated detection techniques to both reduce the cost and improve the overall accuracy. For this to be effective, the species of interest needs to be reliably detected in the target footage, and needs to be tracked to ensure that each animal is only counted once. QUT has already developed an approach that shows promise for koalas, however more research is needed to both improve performance, and model the characteristics of the errors to improve abundance estimation.
The research will build upon an already existing system that uses multiple DCNN models and a simple object tracker to locate instances of a target species in UAV imagery. Main project tasks will include:
- Train and deploy addition DCNN models for detection of animals in complex environments
- Benchmark detector performance, and performance when integrated with object tracking
- Enhance the object tracking to better identify multiple instances of the same animal
- Annotation of additional data (if needed) to further improve detector performance
- Investigate how abundance estimation methods can be incorporated within the detection and tracking framework
The research project aims to:
- Improve the detection performance of the existing system
- Investigate how local appearance information and flight data can be used to improve re-detection and reduce double counting of animals
- Investigate if and how abundance estimation methods could be incorporated into the system
Skills and experience
Candidates must have strong programming experience, with prior experience using either python or c++. Prior experience with machine learning/computer vision is desireable, but not required.
Contact the supervisor for more information.