Tehnički vjesnik, Vol. 32 No. 5, 2025.
Izvorni znanstveni članak
https://doi.org/10.17559/TV-20240708001834
Human Action Recognition Using Explainable Features and Sparse Motion History Images
Wei Yang
; Jiangxi University of Technology, The Center of Collaboration and Innovation, Nanchang, China 330098
Yitong Zhou
; Jiangxi University of Chinese Medicine, The College of Computer Science, Nanchang, China 330004
Jianying Xiong
; Jiangxi University of Chinese Medicine, The College of Computer Science, Nanchang, China 330004
Shiwei Zhang
; The Hanlin Hangyu (Tianjin) Industrial Co., Ltd. Tianjin, China 301899
Lei Zhang
; The Hanlin Hangyu (Tianjin) Industrial Co., Ltd. Tianjin, China 301899
Leiyue Yao
; Jiangxi University of Chinese Medicine, The College of Computer Science, Nanchang, China 330004
*
* Dopisni autor.
Sažetak
This study proposes a novel approach to human action recognition (HAR) via depth sensor data. We introduce explainable features derived from skeleton sequences and a sparse motion history image (SMHI) to effectively capture motion characteristics. Our method addresses the limitations of current approaches by reducing the computational requirements while maintaining high accuracy. We propose a SlowFast-like network that combines these features for efficient HAR. Experiments on three datasets demonstrate the effectiveness of our approach, which achieves competitive accuracy with fewer features. The method also ensures user privacy by relying solely on skeleton data. This research contributes to the theoretical advancement of HAR and its practical application in various fields.
Ključne riječi
data augmentation; human action recognition, motion feature matrix; skeleton-based HAR, video encoding
Hrčak ID:
335046
URI
Datum izdavanja:
30.8.2025.
Posjeta: 710 *