Human Activity Recognition Machine With an Anchor-Based Loss Function

• Published in: IEEE sensors journal 2022-01, Vol.22 (1), p.741-756
• Main Authors: Jin, Lei, Wang, Xiaojuan, Chu, Jiaming, He, Mingshu
• Format: Article
• Language: English
• Subjects: Accuracy; Classification; Clustering; Convolutional neural networks; Deep learning; Euclidean geometry; Face recognition; Feature extraction; Human activity recognition; Machine learning; Measurement; metric learning; Moving object recognition; Networks; open-set classification; Sensors; Training
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2021.3130761

More recently, Human Activity Recognition (HAR) based on sensors has become a hot topic due to its wide application. Researchers significantly reduce the cost of feature extraction and improve the accuracy of recognition by introducing deep learning networks. However, human activity data are greatly affected by the inter-personal variability which brings the interclass similarity and the intraclass diversity. They not only increase the difficulty of the closed-set classification, but also affect the performance on the open-set problem. To solve the problem, we design a framework using a loss function of Euclidean distance and a high-dimensional embedding layer to enhance the ability of deep learning networks to mine discriminative features. Furthermore, we define two kinds of open-set problems in HAR: the pseudo open-set problem and the completely open-set problem. We propose a new clustering method based on the Euclidean distance for the pseudo open-set problem, which reduces the computation cost and improves the accuracy. For the completely open-set problem ignored by other researches, we introduce the rejection score to evaluate the distance score between samples and all known classes, and realize the completely open-set classification. We conduct experiments using four common deep learning networks on three public datasets: OPPORTUNITY, PAMAP2 and UniMiB-SHAR. The results show that the performances of the model modified by our method are much better than those of the original model.