Novel Subject-Dependent Human-Posture Recognition Approach Using Tensor Regression

• Published in: IEEE sensors journal Vol. 25; no. 1; pp. 1041 - 1053
• Main Authors: Yan, Kun, Liu, Guannan, Xie, Rende, Fang, Shih-Hau, Wu, Hsiao-Chun, Yu Chang, Shih, Ma, Li
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.01.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural networks; Codes; Datasets; Deep learning; Dynamic multidimensional time-series segmentation; Feature extraction; human-posture recognition Kinect data; Multidimensional data; One shot learning; Performance evaluation; Regression models; Segments; skeletal graph; Support vector machines; tensor regression; Tensors; Three-dimensional displays; Training; Transformers; Vectors
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2024.3493893

A novel dynamic human-posture recognition approach using tensor regression is proposed in this work. In our proposed approach, a new dynamic segmentation scheme based on hidden logistic regression (HLR) is first undertaken to segment multidimensional skeletal graph data. Within each segment of multidimensional data, a new feature tensor consists of high-dimensional skeletal-graph time-series (SGTS) involving multijoint 3-D coordinates and their temporal differences. Regression models can thus be trained from these collected feature tensors with respect to each type of human posture of interest. Experiments using real-world Kinect data are conducted to evaluate the effectiveness of our proposed novel tensor-based human-posture recognition scheme. In comparison with two prevalent deep learning models, namely the graph convolutional network (GCN) and the Transformer, our proposed novel tensor-based human-posture recognition approach can achieve the highest recognition accuracy of <inline-formula> <tex-math notation="LaTeX">{97}\% </tex-math></inline-formula>. Furthermore, we have evaluated the performance of our proposed new method using the open-source Kinect dataset, namely the UTKinect dataset, for one-shot learning. Our proposed novel tensor-based human-posture recognition approach still significantly outperforms the aforementioned prevalent deep learning models for one-shot learning.