The Distributed Sensing Systems (DSS) Group is currently developing a low-cost LoRaWAN-enabled smart camera device and compression mechanisms to afford frequent images and classifications to be transmitted over the highly constrained communications for high temporal-spatial ground-based monitoring of animal (pest/wildlife) presence. Audio recorders can address the challenges of camera-based monitoring by recording species over far larger areas. Researchers increasingly place audio recorders in forests and other ecosystems to monitor birds, insects, frogs, and other animals. Such audio-based monitoring falls in the bigger area of Bioacoustics, which is the study of the production, transmission and reception of animal sounds. The proposed project aims to take these advantages of audio monitoring to further improve ground-based monitoring of animals in the ongoing DSS project.
Deep learning, a branch of Artificial Intelligence (AI), has revolutionised audio and image processing. This project aims to use deep learning to develop a highly accurate animal species recognition system from the audio recording. Existing studies use deep learning techniques to recognise animal species from audio, but the focus is not on deployability but on improving accuracy metrics. As a result, current state-of-the-art models are far too computationally expensive to deploy using today's resource-constrained edge devices. This project aims to address this challenge enabling the processing of AI algorithms on edge devices (known as Edge AI). The aims are as follows:
Develop a lightweight deep learning audio classifier that (a) distributes classification tasks to multiple edge device and (b) integrate the results to make classification decision.
Propose distributed onboard training techniques subject to the limited bandwidth and limited processing power of edge devices. This can be useful when the classification and data collection need to occur simultaneously. A preloaded model on edge can be updated with the on-device training data without needing bandwidth to transfer the training data into a central location for training. Only any change in the model needs to be sent to the central location for updating the global model.
Propose an onboard lightweight fusion of audio with images to improve the performance of the classification.
Conduct on-device testing and report operating requirements.
Besides ground-based monitoring, some other high-level opportunity this project can link to is the monitoring of environmental health for sensitive ecosystems (reefs, rainforests), especially when looking at recovery after natural disasters (floods, bushfires, coral bleaching events); or when characterising the industrial impacts on the ecosystems (e.g., livestock/mining industries).