Research

Image sequence analysis

Computer vision and machine learning on facial feature detection

Yonghze Yan is a second year PhD Candidate at UCA, ComSee team under supervision of Thierry Chateau, Christophe Garcia, Christophe Blanc and Stefan Duffner. His research interests are in computer vision and machine learning on facial feature detection. His work focuses on the study and development of the solution for estimating and monitoring features real-time 3D body and illuminance for simulation of makeup, hair coloring and accessories.

Weather Classification with traffic surveillance cameras

Road operators are using Intelligent Transport Systems composed of road side cameras for traffic management. The artificial vision algorithms used for automatic detection may be impacted by adverse weather conditions. Therefore, it is necessary to improve these algorithms in such conditions. In addition, the applications developed to operate in road context impose a perfect reliability of operation including in adverse weather conditions. There are many works that allow weather classification but they do not take into consideration all the degraded conditions. In this paper, we propose a method based on convolutional neural networks to classify adverse weather conditionsfrom a road camera. This method could be extended to on-board cameras used by autonomous vehicles. We also present the weather image databases that we use to evaluate our learning.

Ma thèse au CEREMA : analyse d’images par méthode de Deep Learning appliquée au contexte routier en conditions météorologiques dégradées  

2D-3D Fusion for Road Object Detection and Tracking on Autonomous Vehicles

L’objectif de la thèse est l’etude et la conception d’une nouvelle méthode de fusion temps réel des informations de flux vidéos d’une caméra RGB et des nuages de points délivrés par un LiDAR dans l’optique d’obtenir un algorithme de détection et de tracking d’obstacles (mobiles et immobiles) en temps réel pour des véhicules autonomes. Nous prenons le parti durant cette thèse de se focaliser sur les méthodes ‘apprentissage automatique, en particulier les méthodes de deep learning. Les obstacles ayant une grande variabilité, nous nous focalisons sur 3 classes d’objets :
— les 4 roues (voitures, vans, camion, etc.), les usagers les plus présents dans le domaine de l’ego-véhicule. Des modèles décrivant leur dynamique suivant leur état existent.
— les piétons : usagers fragiles et très présents en milieu urbain, lents mais plus imprévisibles.
— les 2-roues (vélos, motos) : usagers fragiles mais pouvant acquérir une grande vitesse.

Technical study of neural networks compression for embedded systems

Convolutional neural networks are the most efficient solutions for image recognition and object tracking.
However, better performances is always followed by more layers in the network which will also increase the computation time.
Recent works try to address this compression problem and to reduce complexity by deleting layers, by decreasing the precision of the weights or by using optimized architectures.
The aim of this work is to suggest compression methods that are optimized for mobile architectures.

3D vision and geometry