AutoMR: A Universal Time Series Motion Recognition Pipeline

Likun Zhang, Sicheng Yang, Zhuo Wang, Haining Liang, Junxiao ShenarXiv 2025 · OpenReview

Abstract

In this paper, we present an end-to-end automated motion recognition (AutoMR) pipeline designed for multimodal datasets. The proposed framework seamlessly integrates data preprocessing, model training, hyperparameter tuning, and evaluation, enabling robust performance across diverse scenarios. Our approach addresses two primary challenges: 1) variability in sensor data formats and parameters across datasets, which traditionally requires task-specific machine learning implementations, and 2) the complexity and time consumption of hyperparameter tuning for optimal model performance. Our library features an all-in-one solution incorporating QuartzNet as the core model, automated hyperparameter tuning, and comprehensive metrics tracking. Extensive experiments demonstrate its effectiveness on 10 diverse datasets, achieving state-of-the-art performance. This work lays a solid foundation for deploying motion-capture solutions across varied real-world applications.

Introduction

Motion recognition is currently a fragmented field. Researchers and engineers face a "Wild West" of data: every dataset comes with different sensor modalities (IMUs, skeletal tracking, sEMG), varying sampling rates, and unique noise profiles. Traditionally, this forces a cycle of manual, task-specific engineering involving custom preprocessing, model selection, and the "black art" of manual hyperparameter tuning. This high barrier to entry makes motion recognition inaccessible to non-specialists in fields like healthcare or robotics and leads to repetitive, computationally expensive development cycles.

Methodology

AutoMR introduces a modular, hierarchical architecture that automates the entire lifecycle of a motion recognition model. The pipeline is designed to bring "plug-and-play" simplicity to time-series motion and gesture recognition.

The AutoMR pipeline architecture.
  • Standardized Data Module: Defines a structured format for modality specifications and sampling parameters, allowing it to ingest diverse datasets into a unified pipeline.
  • Model Factory: The core engine uses QuartzNet, utilizing 1D depthwise separable convolutions. These are highly efficient at capturing temporal patterns in sequential data while maintaining a low parameter count.
  • Automated Hyperparameter Tuning: Instead of manual grid searches, AutoMR employs SMAC (Sequential Model-based Algorithm Configuration). This Bayesian optimization approach iteratively refines the search space for learning rates, dropout, and QuartzNet-specific configurations.
  • Training Module: Built on TorchTNT, the module incorporates advanced features like gradient clipping, anomaly detection, and dynamic learning rate scheduling.

Experimental Setup

The authors evaluated AutoMR on 10 diverse benchmark datasets to validate its universal applicability:

  • IMU-based: MHEALTH, UCI-HAR, and four subsets of the OPPORTUNITY dataset.
  • Skeletal/Motion Capture: SHREC2021, Berkeley-MHAD, and LMDHG.
  • Electromyography (sEMG): DB4 (hand gesture recognition).

This variety tests the model's ability to handle both low-dimensional sensor data and high-dimensional skeletal projections.

Results and Performance Analysis

AutoMR achieved state-of-the-art (SOTA) performance on 8 out of the 10 datasets evaluated. Notably, on the OPPORTUNITY dataset, AutoMR outperformed previous SOTA models by over 5%. Ablation studies further confirmed the stability of the automated pipeline, achieving 99.81% accuracy on MHEALTH and 97.05% on UCI-HAR.

AutoMR performance vs. SOTA benchmarks.

Key insights from the results include:

  • Efficiency via 1D Convolutions: 1D convolutions are often sufficient and significantly faster for motion recognition compared to heavier 2D CNNs or LSTMs.
  • Automation Superiority: Automated tuning via SMAC consistently matched or outperformed manual tuning by experts, suggesting that human intervention is often a bottleneck in optimization.
  • Universal Bottlenecks: The model struggled slightly with highly unstructured or noisy data (e.g., DB4 and LMDHG), highlighting remaining challenges with "messy" real-world signals.

Conclusion

While AutoMR excels at structured sensor data, its current reliance on 1D convolutions may miss complex spatial dependencies better captured by 2D CNNs or Graph Neural Networks. Future work will explore hybrid models and transformer-based architectures to improve handling of long-range dependencies and unsegmented data. Despite these limitations, AutoMR provides a robust, automated foundation for deploying motion-capture solutions across varied real-world applications.

Summary updated

Abstract

In this paper, we present an end-to-end automated motion recognition (AutoMR) pipeline designed for multimodal datasets. The proposed framework seamlessly integrates data preprocessing, model training, hyperparameter tuning, and evaluation, enabling robust performance across diverse scenarios. Our approach addresses two primary challenges: 1) variability in sensor data formats and parameters across datasets, which traditionally requires task-specific machine learning implementations, and 2) the complexity and time consumption of hyperparameter tuning for optimal model performance. Our library features an all-in-one solution incorporating QuartzNet as the core model, automated hyperparameter tuning, and comprehensive metrics tracking. Extensive experiments demonstrate its effectiveness on 10 diverse datasets, achieving state-of-the-art performance. This work lays a solid foundation for deploying motion-capture solutions across varied real-world applications.
BibTeX
@misc{zhang2025automr,
  author = {Likun Zhang and Sicheng Yang and Zhuo Wang and Haining Liang and Junxiao Shen},
  title = {{AutoMR: A Universal Time Series Motion Recognition Pipeline}},
  year = {2025},
  eprint = {2502.15228},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.15228},
}
Figure