Skip to content

MoPFormer: Motion-Primitive Transformer for Wearable-Sensor Activity Recognition

Conference: NeurIPS 2025 arXiv: 2505.20744 Code: Available Area: Interpretability Keywords: Motion Primitives, Transformer, Wearable Sensors, Activity Recognition, Temporal Decomposition

TL;DR

MoPFormer is proposed to decompose wearable sensor signals into sequences of motion primitives and model their temporal dependencies via a Transformer, surpassing state-of-the-art methods on multiple HAR benchmarks while maintaining a lightweight architecture.

Background & Motivation

Background: Human activity recognition (HAR) using wearable sensors is widely applied in health monitoring and sports analysis; mainstream approaches employ CNNs or RNNs to process raw signals directly.

Limitations of Prior Work: (1) Raw signals contain substantial noise, and varying sampling rates hinder generalization; (2) CNNs struggle to capture long-range temporal dependencies; (3) standard Transformers adopt unreasonable tokenization strategies for HAR signals.

Key Challenge: The continuous nature of sensor signals conflicts with the discrete token format required by Transformers.

Key Insight: Human activities can be naturally decomposed into motion primitives (e.g., "raise arm," "take a step"), which serve as more semantically meaningful tokens for the Transformer.

Method

Overall Architecture

Input IMU signals → Motion primitive extraction (learned segmentation + encoding) → Primitive-sequence Transformer → Activity classification.

Key Designs

  1. Motion Primitive Extraction

  2. Function: Automatically decomposes continuous sensor signals into discrete primitive sequences.

  3. Mechanism: A learned segmentation network identifies primitive boundaries in the signal; an encoder maps each segment to a fixed-length primitive embedding.

  4. Design Motivation: Primitives constitute more natural and stable representation units that are invariant to sampling rate variations.

  5. Primitive Transformer

  6. Function: Models temporal dependencies among motion primitives.

  7. Mechanism: Standard Transformer encoder with positional encoding, multi-head self-attention, and feed-forward layers.

  8. Design Motivation: The primitive sequence is far shorter than the raw signal (10–20 primitives vs. hundreds of sample points), yielding high computational efficiency.

  9. Lightweight Design

  10. Function: Keeps the model compact for deployment on edge devices.

  11. Mechanism: Small embedding dimensions (64–128), shallow Transformer (2–4 layers), and parameter sharing.
  12. Design Motivation: Wearable devices have limited computational resources.

Loss & Training

Cross-entropy classification loss combined with an auxiliary primitive segmentation loss. The model is trained end-to-end.

Key Experimental Results

Main Results

Method UCI-HAR Acc↑ PAMAP2 F1↑ Opportunity F1↑ Params↓
DeepConvLSTM 93.2% 89.5% 85.3% 2.1M
InceptionTime 94.1% 90.8% 86.7% 3.5M
HAR-Transformer 94.5% 91.2% 87.1% 4.2M
MoPFormer 95.8% 92.7% 89.3% 0.8M

Ablation Study

Configuration UCI-HAR Acc Note
Fixed-window tokenization 93.8% Standard sliding-window segmentation
Learned segmentation, no Transformer 94.2% Primitives + MLP
Learned segmentation + Transformer 95.8% Full model

Key Findings

  • MoPFormer outperforms all baselines on every benchmark while achieving the smallest parameter count (0.8M vs. 4.2M).
  • Motion primitive decomposition contributes most significantly — even replacing the Transformer with an MLP yields better results than standard methods.
  • Strong cross-sampling-rate generalization — performance drops only 1.2% when transferring from 50 Hz to 25 Hz, compared to 5%+ for standard methods.

Highlights & Insights

  • Naturalness of Motion Primitives: Human motion is inherently composed of elemental primitives, making this representation more aligned with kinematic reality. The approach is transferable to robot action recognition, sign language recognition, and related domains.
  • Efficiency Advantage: Primitive sequences are substantially shorter than raw signals, drastically reducing the computational cost of the Transformer while simultaneously improving accuracy.
  • Robustness: Resilience to sampling rate variation is a critical practical advantage for real-world deployment.

Limitations & Future Work

  • The interpretability of learned primitive segmentations remains to be validated — it is unclear whether the discovered primitives correspond to genuine motion units.
  • Evaluation is limited to IMU data; extension to other sensor modalities (EMG, pressure) is necessary.
  • Primitive decomposition may be inaccurate for complex transitional activities (e.g., "sitting down immediately followed by standing up").
  • vs. HAR-Transformer: Applies fixed-window tokenization directly to raw signals, ignoring the natural structure of motion; MoPFormer's learned segmentation provides a more principled alternative.
  • vs. DeepConvLSTM: A CNN-RNN hybrid with a large parameter count and insufficient capacity to capture long-range dependencies.

Rating

  • Novelty: ⭐⭐⭐⭐ The use of motion primitives as tokens is intuitive and effective.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Three datasets with comprehensive ablation studies.
  • Writing Quality: ⭐⭐⭐⭐ Clear and well-structured presentation.
  • Value: ⭐⭐⭐⭐ High practical value for real-world wearable HAR applications.