Abstract
Behavioral context recognition aims to identify an individual’s situational state and has broad application prospects across multiple fields. With the proliferation of smart devices such as smartphones and wearables, sensor data have become powerful tools for recognizing human behavior. Although deep neural networks have shown strong performance in behavioral context recognition, they typically rely on static inference structures. However, given the inherently multimodal and multi-label nature of this task, such static designs exacerbate computational complexity, leading to excessive computational demands that hinder deployment on resource-constrained devices. To address these issues, we propose EBCR, a computation-efficient framework for multimodal multi-label behavioral context recognition. The framework employs a hierarchical modality aggregation mechanism to reduce redundant multimodal computation and introduces a relation-aware loss function to explicitly model label dependencies, thereby improving both recognition accuracy and the reliability of early-exit decisions. On the Extrasensory and ETRI Lifelong 2020 datasets, EBCR achieves a balanced accuracy of 74.6% and 63.0%, respectively, comparable to state-of-the- art methods. When deployed on a Raspberry Pi 5, EBCR operates with only 19.21–25.95 MFLOPs, achieving 4–5 ms latency and 0.020–0.032 J energy consumption, resulting in 1.1×–2.5× faster inference and 17.6%–66.7% lower energy cost compared to existing models. These results demonstrate the effectiveness of EBCR in achieving accurate and efficient multi-label behavioral context recognition under real-world resource constraints.
Funding Information
  • National Natural Science Foundation of China (62303496)
  • National Natural Science Foundation of China (62573441)
  • Basic and Applied Basic Research Foundation of Guangdong Province (2021A1515110029)
  • Shenzhen Medical Research Fund (D250403003)