Vision for the project
To develop a publicly available dataset and benchmarks of smoke-detecting camera imagery, and;
To develop new computer vision algorithms for ignition detection and bushfire monitoring
Research
Ignition detection algorithms are developed based on footage from cameras mounted on towers at high vantage points over bushland, with the aim of recognising patterns of smoke from fires.
Detection/segmentation models are developed, allowing for fire detection from camera footage. This includes a new state-of-the-art uncertainty estimation, so that users can know when detection is uncertain.
A specialised camera has been mounted on the Mount Stromlo Observatory site (an ANU campus), and the project is also receiving data from several other cameras around Australia and overseas to train algorithms. Algorithms are improved by training with greater volumes of data, and novel video analytics with low rates of false positives are developed.
Research aims to establish a publicly available dataset of smoke-detecting camera imagery. This allows for the aggregation and consolidation of data from a range of different organisations and jurisdictions, to be uploaded in a nationally consistent standardised format.
Data on a publicly available dataset is to be openly available for researchers and industry to use for training AI algorithms for the automatic detection of bushfire ignitions from camera imagery, leading to the development of the best possible algorithms. This is of benefit to the nation, increasing the speed and reliability with which ignitions are detected, and reducing danger to fire crews. It also allows for the validation and evaluation of commercial AI programs.
Chief Investigators
Project Team
Gao Zhu