6533b82dfe1ef96bd1291b80

RESEARCH PRODUCT

Détection automatique de chutes de personnes basée sur des descripteurs spatio-temporels : définition de la méthode, évaluation des performances et implantation temps-réel

subject

description

We propose a supervised approach to detect falls in home environment adapted to location andpoint of view changes. First, we maid publicly available a realistic dataset, acquired in four differentlocations, containing a large number of manual annotation suitable for methods comparison. We alsodefined a new metric, adapted to real-time tasks, allowing to evaluate fall detection performance ina continuous video stream. Then, we build the initial spatio-temporal descriptor named STHF usingseveral combinations of transformations of geometrical features and an automatically optimised setof spatio-temporal descriptors thanks to an automatic feature selection step. We propose a realisticand pragmatic protocol which enables performance to be improved by updating the training in thecurrent location with normal activities records. Finally, we implemented the fall detection in Zynqbasedhardware platform similar to smart camera. An Algorithm-Architecture Adequacy step allowsa good trade-off between performance of classification and processing time