AutoMR: A Universal Time Series Motion Recognition Pipeline

TL;DR

AutoMR addresses fragmentation in motion recognition by delivering an end-to-end automated pipeline for multimodal time-series data. It standardizes preprocessing, unifies model training, and automates hyperparameter tuning using SMAC, with QuartzNet serving as the core sequence model. Evaluations across ten datasets show state-of-the-art performance on eight datasets, including notable gains on OPPORTUNITY, while revealing limitations on highly noisy data such as DB4 and LMDHG. The work outlines a scalable, accessible path for deploying motion capture solutions and provides an open-source resource to foster reproducibility and collaboration.

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.

Figures (5)

• Figure 1: Key challenges in motion recognition: (1) Dataset variability across different sensor types increases preprocessing complexity, (2) Model adaptation requires dataset-specific optimizations, leading to high computational costs, (3) Hyperparameter tuning demands expertise and computational resources, and (4) Deployment barriers limit accessibility for non-specialists. These challenges highlight the need for a scalable and automated framework.
• Figure 2: AutoMR end-to-end architecture, illustrating the complete workflow from data preprocessing to model training and hyperparameter tuning. The framework standardizes diverse datasets, selects optimal model configurations, and ensures efficient training and deployment for motion recognition across different sensor modalities.
• Figure 3: AutoMR's modular and hierarchical architecture, illustrating the interaction between core modules and dataset-specific training scripts. The upper layer consists of fundamental components for dataset preprocessing, augmentation, model selection, training, and hyperparameter tuning, ensuring a standardized and optimized workflow. The lower layer contains dataset-specific execution scripts that utilize these core modules for model training on individual datasets, enabling efficient adaptation across diverse sensor modalities.
• Figure 4: Overall accuracy comparison between AutoMR and SOTA models across ten datasets. AutoMR achieves superior performance on eight datasets, highlighting its effectiveness in generalizing across diverse gesture recognition tasks.
• Figure 5: Ablation study comparing automatic and manual hyperparameter tuning across ten datasets. Results indicate that automatic tuning performs on par with or better than manual tuning in most cases, validating the efficiency and practicality of AutoMR’s optimization strategy.