TOPIC

• Titile: Task-Difficulty Aware Meta-Learning for Adaptive Few-Shot Human Activity Recognition using UWB Sensors

INFORMATION

• Producer: Ji-Sang Park
• Published: IEEE Sensors Conference'25
• Paper: Link

IMAGE AND VIDEO

OVERVIEW

• This study presents the Difficulty-Adaptive Strategy Selector for Learning (DASSL), a task-aware metalearning framework designed for adaptive few-shot human activity recognition (HAR) with ultra-wideband (UWB) sensors. To validate the proposed system, we evaluated DASSL using data acquired with our self-developed UWB sensor. DASSL dynamically adjusts key training parameters-learning rate, regularization, and inner-loop update steps-based on taskspecific statistical difficulty indicators, including intra-class variance, inter-class similarity, and query-support alignment. Experimental results show that DASSL achieves 90.2% accuracy, outperforming fixed-strategy and task-agnostic metalearning baselines by 6.5% and 4.9%, respectively. Furthermore, the adaptive selection of hyperparameters enhances interpretability and computational efficiency, making the framework well-suited for deployment in resourceconstrained elderly health assistance systems.