Section: New Results

Activity Recognition Applied on Health Care Application

Participants : Rim Romdhane, Veronique Joumier, François Brémond.

The aim of this work is to propose a constraint-based approach for video event recognition with probabilistic reasoning for handling uncertainty. This work was validated on health care applications.

Event Recognition

We propose an activity recognition framework which is able to recognize composite events with complex temporal relationships. We consider different aspects of the uncertainty of the recognition during the event modeling and the event recognition process to overcome the noise or missing observations which characterize real world applications.

To reach this goal, we manage the uncertainty in the event modeling and event recognition processes by a combination of logical and probabilistic reasoning for handling uncertainty. We improve the event description language developed in Pulsar team and introduce a new probabilistic description based approach to gain in flexibility for event modeling by adding the notion of utility. Utility expresses the importance of sub-events to the recognition of the whole event. We compute the probability of recognition for both primitive (i.e. elementary) events and composite events based on Bayesian theory.

We compute the probability that the event $e$ is recognized given a sequence of observations $O$ as described in  [48] . The observations consist of the set of the physical objects $p{o}_e$ moving in the scene. If the probability of an event is over a predefined threshold, the event is recognized.

Health Care Application

The proposed event recognition approach is validated using the videos from the health care application SWEETHOME (http://cmrr-nice.fr/sweethome/ ) and CIUSante (https://extranet.chu-nice.fr/ciu-sante ). We have worked in close collaboration with clinicians from Nice hospital to evaluate the behaviours of Alzheimer patients. We have first model 69 event models for health care application using our event modeling formalism. With the help of clinicians we have established a scenario protocol. The scenario is composed of three parts: (1) directed activities (10 min), (2) semi directed activities (20 min), (3) free activities (30 min). Experiments have been performed in a room of Nice hospital equipped with 2 video cameras where 45 elderly volunteers have spent between 15 min to 1 hour.Volunteers include Alzheimer patients, MCI (mild cognitive impairment) and healthy elderly.

The study described in  [38] and  [27] shows the ability of the proposed automatic video activity recognition system to detect activity changes between elderly subjects with and without dementia during a clinical experimentation. A total of 28 volunteers (11 healthy elderly subjects, 17 Alzheimer’s disease patients (AD)) participate to the experimentation. The proposed study shows that we could differentiate the two profiles of participants based on motor activity parameters, such as the duration of the recognized activities, the strike length and the walking speed, computed from the proposed automatic video activity recognition system. These primary results are promising and validating the interest of automatic analysis of video as an objective evaluation tool providing comparative results between participants and over the time.