Subject short description
The objective of this thesis is to develop a deep neural network model capable of predicting the behavior of aquatic animals in novel environments. The model will be trained on a hybrid corpus combining video data and fluid flow data derived from numerical simulations and experimental measurements.
Proposed Methodology:
1) Automated Video Analysis: Detection and segmentation of animals to extract their locomotor and physiological characteristics ;
2) Data Fusion: Alignment of biological parameters with the physical characteristics of the flow ;
3) Multimodal Training: Design of an AI model capable of simultaneously interpreting heterogeneous data sources ;
4) Prediction: Generation of trajectories in new habitats or hydrological contexts ;
5) Evaluation: Comparison between predicted trajectories and observed data.
Scientific Challenges:
Collaborative work between XLIM and MIA (Rabu2026) has demonstrated the superiority of Deep Learning approaches in processing complex videos characterized by turbulence and visual artifacts. Two major challenges have been identified:
1) Extraction of physiological characteristics: Small-amplitude movements (gills, fins, antennae) are difficult to capture. This thesis plans to use motion magnification methods combined with Physics-Informed Neural Networks (PINNs). These networks allow for the coupling of visual data with mathematical swimming models to accurately estimate their parameters.
2) Multimodal AI Architecture: Proposing an architecture capable of effectively merging video, experimental, and simulated data to ensure reliable prediction of aquatic animal behavior across different experimental or simulated configurations.
Subject full description
Host laboratory
Team ASALI/ICONES of XLIM institute
Laboratoire XLIM UMR CNRS 7252
Bât. H1 – SP2MI
11 Bd Marie et Pierre Curie
TSA 41123
86073 Poitiers Cedex 9
https://maps.google.fr/maps?q=11+bd+marie+et+pierre+curie+86360+futuroscope+chasseneuil
Contacts
- Benjamin BRINGIER benjamin.bringier@univ-poitiers.fr
- Benoit TREMBLAIS benoit.tremblais@univ-poitiers.fr
- Renaud PÉTERI renaud.peteri@univ-lr.fr
Recruitment schedule
Candidate profile
- coming from computer science, signal and image processing, or physic, math Master diploma or equivalent
- skills : signal and image processing, computer vision, data analysis, deep learning, python/C++ programming
- stringly motivated
- scientific rigor, methodology and autonomy
