Intitulé de l'offre : PhD candidate (M/W) in computer vision and AI (H/F)
Référence : UMR7020-FREBOU-001
Nombre de Postes : 1
Lieu de travail : LA GARDE
Date de publication : lundi 22 mai 2023
Type de contrat : CDD Doctorant/Contrat doctoral
Durée du contrat : 36 mois
Date de début de la thèse : 1 septembre 2023
Quotité de travail : Temps complet
Rémunération : 2 135,00 € gross monthly
Section(s) CN : Information sciences: processing, integrated hardware-software systems, robots, commands, images, content, interactions, signals and languages
Description du sujet de thèse
The detection and classification of small mobile objects in a video is a particularly active area of research given the current industrial and societal challenges (autonomous vehicles, the fight against asymmetric threats, etc.). In terms of security in particular, the stealthiness of the drone over long distances and its low cost allows malicious operators to jeopardise the security of sensitive installations and people. However, conventional detection methods are hampered in the case of a small mobile object, such as a drone, due to low spatial resolution (the surface of the target is often reduced to a few pixels) as well as to motion blur.
The aim of this thesis is to study an approach that consists in integrating features related to flight dynamics into the identification process. In particular, the objectives are to study and develop different aspects of mobile target detection, tracking and classification using both optical and dynamic features. These different aspects will be studied jointly with the aim of proposing a unified approach. In particular, a multi-model tracking approach will be used, integrating displacement models resulting from learning. The tracking should also make use of the optical characteristics extracted by the classification process. The different information (optical and dynamic) allowing the classification of the target will be integrated in an incremental way using a recurrent model (LSTM or GRU type).
The study will also concern aspects related to the learning of models in this particular context for which there is no annotated database and for which the constitution of such a database remains a lock.
Various approaches are envisaged to circumvent this problem, including
- The use of "fine tuning" of the model, in which the different modules (relating to the extraction of optical characteristics and dynamics) will be pre-trained separately.
- The generation of synthetic data by a global or local approach using a generative neural network model (GAN or VAE).
- The use of unsupervised or weakly supervised learning methods.
Contexte de travail
The thesis will take place in the LIS UMR 7020 laboratory (www.lis-lab.fr) at the University of Toulon. The LIS conducts fundamental and applied research in the fields of computer science, automation, signal and image. It is composed of 20 research teams and structured in 4 poles. The PhD student will be integrated into the SiIM team of the "Signal & Image" pole.
Contraintes et risques
No particular risks