Article Details

Abstract

Wearable human activity recognition faces challenges in edge deployment and continual learning. Existing methods often overlook the domain shift issue during incremental learning, where data distributions vary due to users, devices, or environments. This paper proposes a lightweight domain-aware prototype calibration (LDPC) framework. By freezing the feature extractor, LDPC stabilizes base knowledge and reduces parameters. A domain-aware prototype calibration mechanism is designed to calibrate biased new class prototypes using base class semantic similarities and adaptively fuse new/old domain knowledge based on domain similarity. Experiments on the Opportunity and PAMAP2 datasets demonstrate that the proposed method achieves significant performance improvements in few-shot continual learning scenarios.

Keywords

human activity recognition; lightweight model; domain-aware prototype calibration; incremental learning

References

1. [1] Kirkpatrick J，Pascanu R，Rabinowitz N，et al. Overcoming Catastrophic Forgetting in Neural Networks [J].Proceedings of the National Academy of Sciences，2017，114（13）：3521-3526.
2. [2] Lopez-Paz D，Ranzato M A. Gradient Episodic Memory for Continual Learning [J]. Advances in Neural Information Processing Systems，2017，30：6467-6476.
3. [3] Li Z，Hoiem D. Learning Without Forgetting [J].IEEE Transactions on Pattern Analysis and Machine Intelligence，2017，40（12）：2935-2947.
4. [4] Xie J，Yan S，He X. General Incremental Learning With Domain-Aware Categorical Representations [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition，2022：14351-14360.
5. [5] Agarwal P，Alam M. A Lightweight Deep Learning Model for Human Activity Recognition on Edge Devices [J].Procedia Computer Science，2020，167：2364-2373.
6. [6] Ordóñez F J，Roggen D. Deep Convolutional and LSTM Recurrent Neural Networks for Multimodal Wearable Activity Recognition [J/OL].Sensors，2016，16（1）：115.https://doi.org/10.3390/s16010115.
7. [7] Wu Donghui, Xu Jing, Chen Jibin, etc Human behavior recognition algorithm based on fusion attention mechanism and CNN-LSTM [J]. Science, Technology and Engineering, 2023, 23 (2): 681-689
8. [8] Shavit Y，Klein I. Boosting Inertial-Based Human Activity Recognition With Transformers [J].IEEE Access，2021，9：53540-53547.
9. [9] Bian S，Liu M，Rey V F，et al. TinierHAR: Towards Ultra-Lightweight Deep Learning Models for Efficient Human Activity Recognition on Edge Devices [C]//Proceedings of the 2025 ACM International Symposium on Wearable Computers，2025：163-169.
10. [10] Zhang Jianhao, Han Wenhao, Yin Quanhe, etc Continuous relationship extraction with few samples based on adversarial prompt sample enhancement and prototype learning [J]. Science and Technology Innovation and Application, 2025, 15 (18): 45-52
11. [11] Zhu F，Zhang X Y，Wang C，et al. Prototype Augmentation and Self-Supervision for Incremental Learning [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition，2021：5871-5880.
12. [12] Gong B，Shi Y，Sha F，et al. Geodesic Flow Kernel for Unsupervised Domain Adaptation [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition，2012：2062-2069.
13. [13] Asadi N，Davari M R，Mudur S，et al. Prototype-Sample Relation Distillation: Towards Replay-Free Continual Learning [PP/OL].（2023-03-26）.https://arxiv.org/abs/2303.14771.

Funding

National College Students' Innovation and Entrepreneurship Training Program (202410285058Z); National Natural Science Foundation of China Youth Science Fund Project (62106167)

About the authors

Lu Chunjiao (2004-), female, Han ethnicity, from Wudalianchi, Heilongjiang Province, undergraduate studies, research direction: deep learning, data mining; Corresponding author: Feng Siwei (1989-), male, Han ethnicity, from Suzhou, Jiangsu, associate professor, PhD, research direction: human behavior recognition, continuous learning.