Research projects

Dem@care project

Dem@Care aspires to contribute to the timely diagnosis, assessment, maintenance and promotion of self-independence of people with dementia, by deepening the understanding of how the disease affects their everyday life and behaviour.

It implements a multi-parametric closed-loop remote management solution that affords adaptive feedback to the person with dementia, while at the same time including clinicians into the remote follow-up, enabling them to maintain a comprehensive view of the health status and progress of the affected person.

Within Dem@care, Signal and Image Processing group at IMS and Video Analysis and Indexing group at LaBRI join their efforts as University Bordeaux 1 (UB1) partner. We contribute to the processing of wearable video for the analysis of activities of daily living, following up on the pioneering research done within the IMMED project.

Financial support: European Community’s Seventh Framework Programme (FP7/2007-2013) under grant agreement 288199, 2011-2015.

Project website: http://www.demcare.eu/

IMMED project

(Indexing MultiMEdia data from wearable sensors for diagnostics and treatment of Dementia)

The IMMED Project aims at developping new applications of signal and image processing and multimedia indexing methods in healthcare. The monitoring of various diseases by means of a (semi) automatic analysis of recorded video makes its first steps in the medical sector due to the maturity of video acquisition and storage technologies.

With the ageing of population, dementia cases consequently increases, but the process leading to diagnosis does not allow identifying all people suffering from dementia in the general population. Hence, an early diagnosis of dementia is still a great challenge to prevent insecurity and health worsening in aged people living at home.

Our intention is to develop a new system to assess cognitive decline in the patients’ life in the most comprehensive manner. By the way, we aim at highlighting early instrumental activities or cognitive processes impairments. This would help in people rehabilitation, by offering strategies to maintain their security and autonomy at home.

Within the IMMED project, we have contributed to several aspects:

• Development of a GoPro wearable camera device adapted to clinical tests with patients
• Estimation of the patient location within their environment at room recognition level. Multiple features classifications approaches coupled with semi-supervised learning have been evaluated.
• Tools for 3D localization used in location based event detection with respect to places of interest.
• Graph-word based object recognition algorithms (in collaboration with LaBRI)
• Multimodal activity recognition algorithms based on the Hierarchial HMM formalism (in collaboration with LABRI and IRIT)
• Coordination with clinicians for the evaluation of tools on real world data acquired on patients (in collaboration with INSERM, LABRI, IRIT)

This project is led in close collaboration with technical partners:

• the Video Analysis and Indexing group at LaBRI (activity and multimedia indexing)
• SaMOVA group at IRIT (audio and multimodal indexing)

as well as end-user partners:

• ISPED, CHU Bordeaux/INSERM (age dementia specialists and clinical tests)

Supervision:

• PhD thesis of Vladislavs Dovgalecs (IMS) on wearable video localization (joint supervision with Yannick Berthoumieu)
• PhD thesis of Svebor Karaman (LABRI) on activity and object recognition for wearable video (joint supervision with Jenny Benois-Pineau, LABRI).
• Postdoc position of Hazem Wannous, on wearable camera device design and localization from wearable video.

Collaboration with: LaBRI, IRIT, ISPED CHU Bordeaux/INSERM.

Financial support: ANR Blanc 2009-2012, PhD grant IMS 2008-2011, BQR Université Bordeaux 1 2008-09, PEPS S2TI-CNRS 2007-08.

More details, see the project web page: http://immed.labri.fr/

Vision sensing for micro-UAV

Co-supervision with Audrey Giremus and Yannick Berthoumieu of the PhD thesis of Julien Metge within a CIFRE contract with the company Fly-n-Sense.

Parametric image alignment

Image alignment is a fundamental task of many vision applications. Over the last decades, it has been used to solve numerous problems related to object tracking, image mosaicking, video compression or augmented reality to cite a few. In this work, we propose a unification of direct approaches that register images by using a local gradient based optimization of displaced image difference error measure. This covers classical approaches such as the Lucas-Kanade algorithm, or more recent and efficient approaches such as the ESM algorithm.

Our approach is based on a closed form error functional in the Bidirectionnal composition space. It uses the powerful Lie Group formulation of the transformation parameters to handle the geometrical aspects of image registration. This framework has led us to propose two new approaches: - Bidirectionnal Composition on Lie Group (BCL), - Asymetric Composition on Lie Group (ACL) and several variants of them. New insights on the trade-off between structural bias and noise related variance have been provided. In particular, those methods have been shown to have interesting properties to automatically adapt in situations where the images are corrupted by strong noise of unknown variances.

Current work is in progress to extend the framework and algorithms to multi-image alignement.

Joint work with: Jean-Baptiste Authesserre (PhD defence dec. 2010), Yannick Berthoumieu. Since 2007.

Related publications:

• J.-B. Authesserre, R. Mégret, Y. Berthoumieu. "Automatic Estimation of Asymmetry for gradient-based Alignment of Noisy Images on Lie Group". Pattern Recognition Letters, vol 32, n° 10, pp 1480-1492, juil. 2011. [HAL]
• R. Mégret, J.-B. Authesserre, Y. Berthoumieu. "Bidirectional Composition on Lie Groups for Gradient-based Image Alignment". IEEE Transactions on Image Processing 19(9): 2369-2381, 2010. [Preprint] [DOI]

PLUS project

The objective of PLUS project (Positionnement Laser Uni-Source) was to develop a 6dof real-time precise helmet positionning system. The Signal and Image Processing Group of IMS has contributed to the proposition, study and evaluation of the 3D vision algorithms and software.

Collaboration between: IMS, Thales Avionics, I2S, Novalase.

Labeled by the “Aerospace Valley” and “the Lasers road” competitiveness cluster.

Financed jointly by the French “Front Unique Interministériel” and the Aquitaine Region Council. From 2007 to 2010.

Tracking using color distribution models

Color distribution models are popular for fast object tracking in videos for applications such as videosurveillance, players tracking in sports, or hand and face tracking for human-machine interfaces. In this work, two aspects of object tracking are considered

Extension of color based tracking to inverse approaches, yielding much improved computationnal efficiency Evaluation methodology to assess the robustness and precision of object tracking

Joint work with: Mounia Mikram (PhD defence dec. 2008), Yannick Berthoumieu, Mohamed Najim. From 2004 to 2008.

Related publications:

• R. Mégret, M. Mikram et Y. Berthoumieu. "Inverse Composition for Multi-kernel Tracking", ICVGIP 2006, Madurai, Inde. [PDF]
• M. Mikram, R. Mégret, Y. Berthoumieu. "Evaluation des performances de descripteurs pour le suivi d'objets", GRETSI 2007, Troyes.
• M. Mikram, R. Mégret, Y. Berthoumieu. "Evaluating Descriptors Performances for Object Tracking on Natural Video Data", ACIVS 2007, Delft, Pays-Bas. [PDF]
• M. Mikram, R. Mégret, M. Donias et Y. Berthoumieu. "Multi-Scale Histograms For Kernel-Based Object Tracking", International Symposium on Communications, Control and Signal Processing (ISCCSP) 2006, Marrakech, Maroc.

Cheeger-cut based image segmentation

Motivated by the recent advances in spectral clustering based on the relation between the non linear p- Laplacian graph operator and the Cheeger cut problem, we propose in this work to study this approach in the context of image segmentation. Based on a l1 relaxation of the graph clustering problem, we show that these methods can outperform usual well known graph based approaches, e.g., min cut/max flow algorithm or l2 spectral clustering, for unsupervised or interactive image segmentation. Experimental results show the benefits of the proposed methodology especially for noisy images or when very few pixels are labeled (less than x pixels) for interactive image segmentation.

Joint work with Ludovic Paulhac and Vinh-Ta (LABRI).

• L. Paulhac, V.-T. Ta and R. Mégret, “Relaxed Cheeger Cut for Image Segmentation”, Proc. of International Conference on Pattern Recognition, Tsukuba, Japon, 2012. [HAL]

Role of visual motion for human-machine interaction

Joint work with: Pierre Salom (PhD defence march 2007), Marc Donias, Yannick Berthoumieu. From 2004 to 2007.

Related publications:

• P. Salom, R. Megret, M. Donias, Y. Berthoumieu. "Dynamic picking system for 3D seismic data: design and evaluation". International Journal of Human-Computer Studies , pp 423-430 (2009).
• P. Salom, R. Mégret et Y. Berthoumieu. "Coupe projective pour une tâche trajectorielle dynamique". Conférence Francophone sur l'Interaction Homme-Machine IHM'05, Septembre 27-30, 2005, Toulouse. p 251-254.
• P. Salom, J. Becerra, M. Donias et R. Mégret. "Projective slice for a dynamic steering task". ACM Symposium on Virtual Reality Software and Technology (VRST), November 7 - 9, 2005, Monterey, USA.