(paper source) Prototype-Guided Physics-Modulated Perceiver for Human Activity Recognition
This repository implements the methodology proposed in the paper "Prototype-Guided Physics-Modulated Perceiver for Human Activity Recognition".
The P2-Perceiver model is designed for sensor data processing, incorporating structure-aware embeddings and a Perceiver backbone for efficient feature extraction from IMU signals like accelerometers, gyroscopes, and magnetometers. It begins with an ELK layer that groups sensors by axes or modalities and applies Expanded Large Kernel (ELK) blocks to generate embeddings, followed by a mixer for fusion. The backbone uses a Perceiver architecture with time-aware cross-attention and self-attention blocks to handle variable-length sequences, projecting inputs to latents and processing them through multiple layers. The overall loss combines cross-entropy for classification with an orthogonal loss on group features to encourage diversity and independence among sensor groups. Additionally, a physics-guided loss is incorporated, including gravity-gyro consistency, accelerometer-gyro correlation, and jerk minimization to enforce physical priors. This design makes the model robust to noise by leveraging structured embeddings and physics constraints, ensuring reliable performance in noisy real-world sensor environments.
If you use this code in your research, please cite:
@article{P2Perceiver,
title = {Prototype-Guided Physics-Modulated Perceiver for Human Activity Recognition},
author={Gyuyeon Lim and Myung-Kyu Yi}
journal={},
volume={},
Issue={},
pages={},
year={}
publisher={}
}
For questions or issues, please contact:
- Gyuyeon Lim : lky473736@gmail.com
This project is licensed under the MIT License - see the LICENSE file for details.