Reconnaissance d’actions à l’aide d’une approche hiérarchique basée sur l’apprentissage machine

Abstract

Nowadays, action recognition’s literature proposes methods performing relatively well on simple actions, but are typically computationally heavy and less suited for actions with inconsistent sequences of execution. Hence, the method proposed in the article attached to this master’s thesis addresses these difficulties. More specifically, it proposes a hierarchical learning approach that consists in recognizing the signature of some time-localized dynamics. Explicitly, these signatures correspond to the maximum and minimum activation of convolution kernels learned throughout the process. While being computationally cheap and robust to redundant/unnecessary information, the proposed method also achieves near state-of-the-art performances in terms of classification accuracy when applied on data provided by wearable inertial sensors.

Date	17 Dec 2020
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Rita Noumeir (Supervisor)