Biodiversity monitoring is a key challenge for the preservation of natural habitats and for low-impact urban planning. This monitoring often requires conducting an inventory of the species present. Methods for automatically detecting species from audio recordings (Passive Acoustic Monitoring) are being developed but face challenges in field deployment. This limitation stems from several factors, including (i) the models’ lack of robustness regarding the distance between the animal and the microphone, and (ii) the absence of a confidence score for the model’s detection. This thesis aims to address these two limitations by (1) recording acoustic data under real-world conditions with distance information, (2) using this data to estimate an animal’s distance, and (3) incorporating distance into a confidence score for automatic detection models, within a context of computational efficiency.
Supervisors: Théo Mariotte (LIUM), Manuel Melon (LAUM), Marie Tahon (LIUM, dir) Candidate profile: MsC. in acoustics with strong interests in machine learning and deep learning or MsC. in machine learning with interests in acoustics and audio signal processing
