A New Residual Neural Network Architecture Using Skeleton Data for Human Activity Recognition
Résumé
Deep learning method has been efficiently applied to Human Activity Recognition in recent years. There are several types of neural network architectures such as convolutional, recurrent with long short term memory, etc. that have been further studied. The recent researches have shown their capability of learning features automatically from videos. However, many challenges still remain open. The key to let machine learn spatio-temporal information remains to be discovered. This paper proposes an original architecture of residual neural network using skeleton data. A new complex linear layer aimed at retrieving the internal relationships of neural cell will be discussed. Our method has been applied on some well-known public data-sets, mainly UTKinect-Action3D and Florence 3D actions. Combining with temporal pre-processing of human activity video, it achieves 99.25% and 99.3% accuracy respectively on UTKinect-Action3D and Florence 3D actions data-set, over-performing by 2% the best state-of-the-art results.
Origine | Fichiers produits par l'(les) auteur(s) |
---|