An Algorithm for On-Sensor Agnostic Detection of Changes in Human Activity for Ultra-Low-Power Applications

1. Problem Statement

Wearable devices utilizing Inertial Measurement Units (IMUs) for Human Activity Recognition (HAR) face a critical energy bottleneck. Current state-of-the-art solutions typically run HAR networks continuously on every data window, regardless of whether the user's activity has changed.

• Inefficiency: Human activity is often repetitive and static for long periods. Continuous classification during these stable phases wastes significant battery life and computational resources.
• Limitations of Existing Solutions:
  • Model Optimization: Techniques like quantization and pruning reduce model size but still require continuous inference.
  • Hierarchical/Early-Exit: Methods like Dendron or early-exit strategies lower cost per window but do not eliminate unnecessary activations when no change occurs.
  • Gating Approaches: Existing gating methods often require offline training on specific datasets or rely on rigid statistical parameters (e.g., CUSUM) that lack generalization across different devices and activity types.

2. Methodology

The authors propose a lightweight, non-parametric change-detection gate that runs continuously on the sensor. This gate acts as a filter, invoking the full, computationally expensive HAR network only when a change in activity is detected.

Core Algorithm Phases

The algorithm operates in real-time with a 3-second sliding window (50% overlap) and consists of four phases:

1. Feature Extraction (Sample-wise):

   • Instead of buffering raw data, features are computed sample-by-sample to minimize memory and latency.
   • 12 Features: Extracted from acceleration and angular velocity, including Euclidean norms, first-order derivatives, gravity components (Euler angles), mean-crossing rates, and peak-to-peak amplitudes.

2. Reference Template Creation:

   • The algorithm identifies the two most informative features for the current activity by analyzing Variance (sustained amplitude) and Mean of Squared Successive Differences (MSSD) (short-term variability/periodicity).
   • These two features are normalized and accumulated into a 10×10 2D histogram (a "template") representing the joint density of the feature pair. This serves as the Reference Template ($T_r$).

3. Current Template Creation:

   • For every new window, a Current Template ($T_n$) is generated using the same two selected features and the same normalization bounds (min/max) established during the reference phase.

4. Comparison and Decision:

   • The similarity between $T_n$ and $T_r$ is calculated using the Normalized Correlation Coefficient (NCC).
   • Decision Logic:
     • If $NCC \geq Threshold$: Activity is stable. No action is taken; the HAR network remains off.
     • If $NCC < Threshold$: A change is detected. The gate triggers the full HAR network, updates the Reference Template, and resets the monitoring cycle.

Calibration

• No Offline Training: The algorithm does not require training on large datasets.
• Device-Specific Tuning: A brief calibration phase (2–3 minutes of data with 8–10 transitions) is used to determine the optimal NCC threshold ($Th_{NCC}$). This threshold is tuned to maximize True Positive Rate (TPR) to ensure no genuine transitions are missed.

3. Key Contributions

• Agnostic & Non-Parametric: The method requires no prior definition of activity classes and makes no assumptions about the statistical distribution of sensor signals.
• Ultra-Low Resource Footprint:
  • Memory: Requires only 4.53 kB (storing feature values and histograms).
  • Compute: Costs approximately 16 kFLOPs per step.
  • Comparison: This is >50× cheaper than typical HAR networks (e.g., Dendron at ~853 kFLOPs).
• High Sensitivity: Designed to prioritize detecting all changes (high TPR) over avoiding false alarms, ensuring the HAR network is never "asleep" during a transition.
• Hardware Feasibility: The low memory and compute requirements allow for implementation directly on Microcontroller Units (MCUs) and smart sensor units without needing a host processor.

4. Experimental Results

The algorithm was evaluated on two public datasets across three device types: Smart Glasses (UCA-EHAR), Smartwatch (WISDM), and Smartphone (WISDM).

• Comparison with Baselines:
  • CUSUM: Achieved high TPR (91–94%) but poor generalization across different device types and transition dynamics.
  • Supervised ML (DT, SVM): Showed inconsistent performance across devices. When trained on unbalanced data (realistic scenario), TPR dropped significantly (e.g., 61% for SVM on smartphone). They also required extensive offline training.
• Power Savings:
  • In realistic monitoring scenarios (UCA-EHAR), the gate reduced the number of windows requiring full HAR inference by 70% (triggering only ~30% of the time).
  • This translates to an estimated 67% reduction in total computational load (FLOPs), significantly extending battery life.

5. Significance and Conclusion

This paper presents a paradigm shift for edge HAR by moving from "always-on classification" to "change-driven activation."

• Practical Impact: By decoupling the continuous sensing from the heavy classification task, the method enables ultra-low-power wearables that can operate for days or weeks on small batteries.
• Generalization: The algorithm's ability to adapt to different sensors (glasses, watches, phones) with only a brief calibration makes it highly scalable for commercial deployment.
• Future Work: The authors plan to validate the system on physical devices in real-time, real-world monitoring conditions to further verify power savings and robustness.

In summary, the proposed "On-Sensor Agnostic" gate offers a robust, lightweight, and highly sensitive solution to the energy bottleneck in wearable HAR, achieving a favorable trade-off between detection reliability and computational cost.