Agile human activity recognition for wearable devices based on online incremental learning.

Area of Science:

• Edge Computing
• Machine Learning
• Sensor Data Analysis

Background:

• Human activity recognition (HAR) on resource-constrained edge devices faces challenges in precision, latency, and adaptation.
• Existing HAR methods often optimize single aspects (e.g., online learning, model sparsification) but lack synergistic optimization for accuracy, latency, and power.
• Non-stationary sensor data streams complicate dynamic balancing of HAR performance metrics.

Purpose of the Study:

• To develop an end-to-end closed-loop adaptive learning framework for HAR on edge devices.
• To synergistically optimize feature extraction, model complexity, and adaptation speed for HAR systems.
• To address the limitations of existing HAR approaches in balancing accuracy, latency, and power consumption.

Main Methods:

• Designed an end-to-end closed-loop adaptive learning framework for HAR.
• Utilized fast principal component analysis for adaptive feature dimensionality reduction.
• Implemented an information theory-based dynamic sparse subnetwork activation for model selection.
• Integrated a low-complexity online incremental learning module for concept drift tracking.

Main Results:

• The framework achieved high accuracies ranging from 85.6% to 97.4% across five datasets.
• Inference latency was approximately 1.0 ms, meeting real-time requirements.
• Demonstrated joint dynamic optimization of feature extraction, model complexity, and adaptation speed.

Conclusions:

• The proposed framework effectively addresses the challenges of HAR on edge devices.
• The system-level synergistic design enables dynamic balancing of accuracy, latency, and power consumption.
• The framework meets real-time performance demands for adaptive human activity recognition.