Scholarlink

<p>Detecting fine-grained micro activities from wearable sensor data is difficult because these subtle motions are often hidden within dominant macro-level activity patterns. Although recent deep learning methods have improved recognition performance, they usually rely on heavy models and high computational cost, which limits their direct use on resource-constrained wearable devices. In this paper, we present a dual-teacher knowledge distillation framework for micro activity recognition using accelerometer signals. The student model is intentionally designed to remain lightweight and efficient, while two teachers provide complementary guidance. The first teacher captures macro-level motion structure, and the second teacher is specifically trained to recognize micro activities. The micro teacher transfers both its output predictions and internal feature representations to the student. To address temporal inconsistencies and representation differences between the teachers and the student, we introduce Tucker Feature Trajectory Alignment (TFTA). This approach projects LSTM-based feature sequences into a shared low-rank space, enabling effective feature alignment over time. The influence of each teacher is adjusted through an adaptive weighting strategy based on a hybrid of Harris Hawks Optimization and the Arithmetic Optimization Algorithm. Since this process relies on population-based optimization, it increases the computational load during training and benefits from GPU-accelerated and parallel processing. We evaluate the proposed framework on a real-world cooking activity dataset, where micro activities occur frequently and are often ambiguous. The method achieves 90.71% accuracy in micro activity recognition while maintaining low inference latency. To examine generalization, we also conduct experiments on the Skoda Mini Checkpoint dataset, an industrial activity recognition benchmark with different motion characteristics. In this setting, we keep the dual-teacher distillation framework and adjust the supervision strategy based on the target dataset. Feature-based distillation is used only when the teacher and the student have compatible representations, while response-based supervision is used to transfer decision-level knowledge. This allows us to evaluate the generalization ability of the proposed framework across different domains. All computationally intensive operations are confined to the offline training phase. As a result, the distilled student model can operate efficiently in real time on wearable and edge devices.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

From macro–micro to micro activity recognition using knowledge distillation paradigm based on accelerometer

  • 期刊论文

  • 发表日期：

    2026年2月10日

查看全文
  • Mitra Pourhasan,
  • Abdolhossein Fathi

摘要

Detecting fine-grained micro activities from wearable sensor data is difficult because these subtle motions are often hidden within dominant macro-level activity patterns. Although recent deep learning methods have improved recognition performance, they usually rely on heavy models and high computational cost, which limits their direct use on resource-constrained wearable devices. In this paper, we present a dual-teacher knowledge distillation framework for micro activity recognition using accelerometer signals. The student model is intentionally designed to remain lightweight and efficient, while two teachers provide complementary guidance. The first teacher captures macro-level motion structure, and the second teacher is specifically trained to recognize micro activities. The micro teacher transfers both its output predictions and internal feature representations to the student. To address temporal inconsistencies and representation differences between the teachers and the student, we introduce Tucker Feature Trajectory Alignment (TFTA). This approach projects LSTM-based feature sequences into a shared low-rank space, enabling effective feature alignment over time. The influence of each teacher is adjusted through an adaptive weighting strategy based on a hybrid of Harris Hawks Optimization and the Arithmetic Optimization Algorithm. Since this process relies on population-based optimization, it increases the computational load during training and benefits from GPU-accelerated and parallel processing. We evaluate the proposed framework on a real-world cooking activity dataset, where micro activities occur frequently and are often ambiguous. The method achieves 90.71% accuracy in micro activity recognition while maintaining low inference latency. To examine generalization, we also conduct experiments on the Skoda Mini Checkpoint dataset, an industrial activity recognition benchmark with different motion characteristics. In this setting, we keep the dual-teacher distillation framework and adjust the supervision strategy based on the target dataset. Feature-based distillation is used only when the teacher and the student have compatible representations, while response-based supervision is used to transfer decision-level knowledge. This allows us to evaluate the generalization ability of the proposed framework across different domains. All computationally intensive operations are confined to the offline training phase. As a result, the distilled student model can operate efficiently in real time on wearable and edge devices.