Szczegóły publikacji
Opis bibliograficzny
Ensemble of multi-channel CNNs for multi-class time-series classification : depth-based human activity recognition / Jacek TRELIŃSKI, Bogdan KWOLEK // W: Intelligent Information and Database Systems : 12th Asian Conference, ACIIDS 2020 : Phuket, Thailand, March 23–26, 2020 : proceedings, Pt. 1 / eds. Ngoc Thanh Nguyen, [et al.]. — Cham : Springer, cop. 2020. — (Lecture Notes in Computer Science ; ISSN 0302-9743 ; LNCS 12033. Lecture Notes in Artificial Intelligence). — ISBN: 978-3-030-41963-9; e-ISBN: 978-3-030-41964-6. — S. 455–466. — Bibliogr. s. 465–466, Abstr. — Publikacja dostępna online od: 2020-03-04
Autorzy (2)
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 129383 |
|---|---|
| Data dodania do BaDAP | 2020-12-29 |
| Tekst źródłowy | URL |
| DOI | 10.1007/978-3-030-41964-6_39 |
| Rok publikacji | 2020 |
| Typ publikacji | materiały konferencyjne (aut.) |
| Otwarty dostęp |  |
| Wydawca | Springer |
| Konferencja | Asian Conference on Intelligent Information and Database Systems 2020 |
| Czasopismo/seria | Lecture Notes in Computer Science |
Abstract
In this work we present a new algorithm for multivariate time-series classification. On multivariate time-series of features we train multi-class, multi-channel CNNs to model sequential data. The multi-channel CNNs are trained on time-series drawn with replacement from a pool of augmented time-series. The features extracted by such bagging meta-estimators are used to train SVM classifiers focusing on hard samples that are close to the decision boundary and multi-class logistic regression classifiers returning well calibrated predictions by default. The recognition is done by a soft voting-based ensemble, built on SVM and logistic regression classifiers. We demonstrate that despite limited amount of training data, it is possible to learn sequential features with highly discriminative power. The time-series were extracted in tasks including classification of human actions on depth maps only. The experimental results demonstrate that on MSR-Action3D dataset the proposed algorithm outperforms state-of-the-art depth-based algorithms and attains promising results on UTD-MHAD dataset. © 2020, Springer Nature Switzerland AG.
