Feature Anchors for Time-Series Sensor-Based Human Activity Recognition

2026-04-27

AI Summary

This work addresses the challenge in human activity recognition (HAR) on wearable devices of simultaneously achieving interpretability and adaptability in feature representation: handcrafted temporal features offer clear semantics but are rigid, whereas deep learning features are flexible yet lack transparency. To bridge this gap, we propose TCNet, which uniquely integrates handcrafted temporal features as explicit “feature anchors” within the network architecture. TCNet dynamically modulates the scale, bias, and gating parameters of these anchors by leveraging time-frequency contextual information extracted from raw IMU signals via neural networks, yielding representations that are both semantically interpretable and task-adaptive. This approach transcends the limitations of conventional preprocessing pipelines and purely end-to-end learning, achieving state-of-the-art performance across five HAR benchmarks—with a peak mF1 score of 94.5% on PAMAP2, surpassing rTsfNet by up to 14.6 percentage points. Ablation studies confirm the efficacy of the anchor-guided mechanism.

Abstract

Wearable Human Activity Recognition (HAR) still lacks a representation that is both explicit and adaptable. Handcrafted time-series features (TSFs) capture meaningful motion statistics and remain competitive on standard benchmarks, but they are usually used as fixed preprocessing outputs. Deep models learn adaptable representations directly from raw signals, but those representations are typically latent and difficult to inspect. We address this gap by treating handcrafted TSFs as feature anchors: explicit intermediate representations that remain inside the model and are adjusted by neural context instead of being discarded. We propose the Temporal Conditioning Network for Feature Anchors (TCNet), which extracts handcrafted anchors, encodes complementary time-domain and frequency-domain context from raw IMU windows, and predicts context-conditioned scale, bias, and gating parameters to modulate anchor groups directly in feature space. This design keeps anchor semantics visible while allowing the representation to adapt to the classification objective. Across five HAR benchmarks, TCNet achieves 70.2% mF1 on USC-HAD, 85.1% mF1 on Daphnet, 93.9% mF1 on MHealth, and 94.5% mF1 on PAMAP2. Relative to rTsfNet, it improves by 4.5 points on USC-HAD, 14.6 points on Daphnet, and 6.5 points on MHealth. Ablations show that the gains come primarily from anchor guidance rather than simple branch fusion, and feature-space analyses indicate that several discriminative TSF families are not reliably accessible in standard latent representations. These results suggest that, for HAR, handcrafted TSFs are most useful when they remain explicit and adaptable within the model. The code is available at: https://github.com/ni-x-lab/TCNet-har

Problem

• Human Activity Recognition
• Time-Series Features
• Feature Representation
• Wearable Sensors
• Interpretable Deep Learning

Innovation

• feature anchors
• time-series features
• human activity recognition
• temporal conditioning
• interpretable representation