IMMED

This pluri-disciplinary project aims at developing and validating a very innovative technology based on wearable camera. It will help defining new methodologies for the diagnosis of dementia. A new complete system will be developed that includes wearable recording and indexing tools of audio/video streams, which will be integrated into a single browsing interface aimed at physicians. Three objectives can be identified: 1) design of a new acquisition ergonomic device for wearable audio and video capture 2) resolving technically challenging issues in multimedia indexing, about automatic video and audio analysis 3) validating these technologies in order to integrate them into the context of clinical studies of dementia It must be emphasized that the objectives are ambitious from the point of view of automatic analysis and indexing of multimedia content. Indeed these contents have a great complexity from the point of view of the variability of the audio and visual events, in a natural environment. 1) Technical objectives The preliminary study that was done in the framework of the PEPS CNRS 2007-08 project yielded a first prototype, that allowed to record the daily activity of a patient during several hours. This device is the first wearable device to acquire data at video frame rate in the context of dementia medical diagnostic. During this project, the design of this prototype will be improved in order to make it simpler and more robust to use for a larger scale deployment, while retaining an identical or improved signal quality. The evaluation of the device will be done in the medical context, by measuring the possibility for the medical staff to use it without risk for the patient and themselves, while producing reproducible data which will be useful for the clinical diagnosis. 2) Solving challenging issues in multimedia indexing In the context of Information Technologies, this project raises true scientific questions. First very large amount of data are to be processed, with a 25 frames/s frame rate during several hours. Second, these videos are very noisy from the point of view of motion and therefore represent a unique source of data for the study of methods than can be employed in a vast field where video data suffers from arbitrary movements, motion blur and where little a priori information is known about the environment. Indeed, the issues raised by video indexing for the observation of patients is very broad and covers major needs expressed in the User Generated Content (UGC) application field. New algorithms for the detection of zones of patient activity within the video will be designed with this respect. Based on these algorithmic advances, a video browsing interface will be developed to allow the patricians to browse through the acquired data in a standard (temporal) way, or in a conceptual way, thanks to the automatic detection of events of interest. This will help them elaborate a diagnosis. 3) Advances for the diagnosis of dementia This technology will be used to document in real life 'ecological' situations the difficulties encountered by the patients in their daily activities. To our knowledge, this is a very original approach in the Alzheimer disease and related syndromes. A fine grained analysis of these difficulties combined with neuropsychology should help in proposing better solutions for their correction. It is also planned to capture an extended video corpus on several patients, to annotate them in terms of events of interest, and to create new methodologies that will be related to this new observation tool. The automatic indexing algorithms proposed by the technological partners in this project will assist the patricians in this task.