Inception inspired CNN-GRU hybrid network for human activity recognition

• Published in: Multimedia tools and applications Vol. 82; no. 4; pp. 5369 - 5403
• Main Authors: Dua, Nidhi, Singh, Shiva Nand, Semwal, Vijay Bhaskar, Challa, Sravan Kumar
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.02.2023; Springer Nature B.V
• Subjects: 1178: Pattern Recognition for Adaptive User Interfaces; Artificial neural networks; Classifiers; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Human activity recognition; Human-computer interface; Multimedia Information Systems; Multivariate analysis; Sensors; Special Purpose and Application-Based Systems; User interfaces; Gated recurrent unit; Convolutional neural network; HAR; Human-computer interaction; Inception; Pattern recognition; Wearable sensors
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-021-11885-x

Human Activity Recognition (HAR) involves the recognition of human activities using sensor data. Most of the techniques for HAR involve hand-crafted features and hence demand a good amount of human intervention. Moreover, the activity data obtained from sensors are highly imbalanced and hence demand a robust classifier design. In this paper, a novel classifier “ICGNet” is proposed for HAR, which is a hybrid of Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU). The CNN block used in the proposed network derives its inspiration from the famous Inception module. It uses multiple-sized convolutional filters simultaneously over the input and thus can capture the information in the data at multiple scales. These multi-sized filters introduced at the same level in the convolution network helps to compute more abstract features for local patches of data. It also makes use of 1 × 1 convolution to pool the input across channel dimension, and the intuition behind it is that it helps the model extract the valuable information hidden across the channels. The proposed ICGNet leverages the strengths of CNN and GRU and hence can capture local features and long-term dependencies in the multivariate time series data. It is an end-to-end model for HAR that can process raw data captured from wearable sensors without using any manual feature engineering. Integrating the adaptive user interfaces, the proposed HAR system can be applied to Human-Computer Interaction (HCI) fields such as interactive games, robot learning, health monitoring, and pattern-based surveillance. The overall accuracies achieved on two benchmark datasets viz. MHEALTH and PAMAP2 are 99.25% and 97.64%, respectively. The results indicate that the proposed network outperformed the similar architectures proposed for HAR in the literature.