Scholarlink

<p>Motion sensor-based human activity detection emerges an interesting field of research because it has widespread real-time usages. The major challenges of real-time recognition systems include the demand of enrich feature extraction capability to address data variability and similarity, a mechanism to select important data that contribute significantly to distinguishing activities, generalisation and scalable ability to handle unseen massive real-world data and computational efficiency for direct use in resource-constrained devices. To balance accuracy and efficiency, we develop a unique resource efficient deep learning framework. With a focus on hybrid deep model, adding lightweight residual long short term memory block, lightweight residual convolutional block and attention mechanism significantly enhances spatio-temporal feature extraction performance without increasing the model’s size. To verify the impact of residual-based feature map processing formulation and attention, we conducted an ablation study by constructing traditional model similar in architecture to residual model under strictly controlled parameter budget, both are examined without and with the presence of attention results in four different model configuration. Experiments are conducted on one processed and two raw public datasets: UCI-HAR handcrafted feature, UCI-HAR raw, and MHEALTH. The recognition success is measured by standard performance metrics and the resource usage is measured using complexity metrics. Our resource efficient residual model with attention shows significant improvement in recognition of activities, achieving an average F1 score of 0.90 or above on 17 out of total 25 activities across three datasets. We further validate the robustness of our architecture on WISDM dataset. With only 15&#xa0;k parameters on the three raw datasets, it obtains 90–91% mean test accuracies, and with just 50&#xa0;k parameters on the one processed dataset, it achieves 96% mean test accuracy. To validate the practical applicability of our original resource efficient model, we conducted deployment experiment on TensorFlow Lite framework reported 100.16&#xa0;KB model size, 4.30&#xa0;ms latency and 23.51&#xa0;mAh energy consumption. The promising results of our model with lower computational requirements making it more appropriate for real-time training and inference on battery-powered wearable 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

Resource efficient, light residual deep attention model for motion sensor-based human activity recognition in IoT application

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • B. Aarthy,
  • G. Tamilpavai

摘要

Motion sensor-based human activity detection emerges an interesting field of research because it has widespread real-time usages. The major challenges of real-time recognition systems include the demand of enrich feature extraction capability to address data variability and similarity, a mechanism to select important data that contribute significantly to distinguishing activities, generalisation and scalable ability to handle unseen massive real-world data and computational efficiency for direct use in resource-constrained devices. To balance accuracy and efficiency, we develop a unique resource efficient deep learning framework. With a focus on hybrid deep model, adding lightweight residual long short term memory block, lightweight residual convolutional block and attention mechanism significantly enhances spatio-temporal feature extraction performance without increasing the model’s size. To verify the impact of residual-based feature map processing formulation and attention, we conducted an ablation study by constructing traditional model similar in architecture to residual model under strictly controlled parameter budget, both are examined without and with the presence of attention results in four different model configuration. Experiments are conducted on one processed and two raw public datasets: UCI-HAR handcrafted feature, UCI-HAR raw, and MHEALTH. The recognition success is measured by standard performance metrics and the resource usage is measured using complexity metrics. Our resource efficient residual model with attention shows significant improvement in recognition of activities, achieving an average F1 score of 0.90 or above on 17 out of total 25 activities across three datasets. We further validate the robustness of our architecture on WISDM dataset. With only 15 k parameters on the three raw datasets, it obtains 90–91% mean test accuracies, and with just 50 k parameters on the one processed dataset, it achieves 96% mean test accuracy. To validate the practical applicability of our original resource efficient model, we conducted deployment experiment on TensorFlow Lite framework reported 100.16 KB model size, 4.30 ms latency and 23.51 mAh energy consumption. The promising results of our model with lower computational requirements making it more appropriate for real-time training and inference on battery-powered wearable edge devices.