Scholarlink

<p>Human activity recognition (HAR) has been a long sought-after field of research, implementing various methods for the same to achieve results with greater accuracy. In this paper, we have worked on one such method for HAR using an ensemble learning approach implemented on datasets consisting of sensor data sequences obtained from different wearable devices like, smartphones and smartwatches. Our overall prediction framework consists of three base classifiers, outputs of which are ensembled by the Sugeno fuzzy integral method. The three base classifiers implemented are: a CNN-LSTM model, a model based on VGG-16 architecture, and a model based on ResNet-18 architecture. In this paper, we see how significantly accurate the model classifies, giving accurate measures of our activities. The framework has obtained an accuracy of 98.28%, 98.4% and 99.72% when evaluated on the three publicly available standard datasets, namely WISDM, PAMAPv2 and MHEALTH, respectively. The results are better than the state-of-the-art accuracies found in the recent literature. The source code is available at <a href="https://github.com/apusarkar2195/FuzzyEnsemble">https://github.com/apusarkar2195/FuzzyEnsemble</a>.</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

Sugeno fuzzy integral based ensemble of deep learners for human activity recognition using sensor data

  • 期刊论文

  • 发表日期：

    2025年6月26日

查看全文
  • Pushan Mondal,
  • Apu Sarkar,
  • Ram Sarkar

摘要

Human activity recognition (HAR) has been a long sought-after field of research, implementing various methods for the same to achieve results with greater accuracy. In this paper, we have worked on one such method for HAR using an ensemble learning approach implemented on datasets consisting of sensor data sequences obtained from different wearable devices like, smartphones and smartwatches. Our overall prediction framework consists of three base classifiers, outputs of which are ensembled by the Sugeno fuzzy integral method. The three base classifiers implemented are: a CNN-LSTM model, a model based on VGG-16 architecture, and a model based on ResNet-18 architecture. In this paper, we see how significantly accurate the model classifies, giving accurate measures of our activities. The framework has obtained an accuracy of 98.28%, 98.4% and 99.72% when evaluated on the three publicly available standard datasets, namely WISDM, PAMAPv2 and MHEALTH, respectively. The results are better than the state-of-the-art accuracies found in the recent literature. The source code is available at https://github.com/apusarkar2195/FuzzyEnsemble.