基于融合Capsule-Transformer的人体活动识别模型

Human Activity Recognition Model Based on the Fusion of Capsule-Transformer

摘要: 近年来，基于可穿戴传感器的人体活动识别在智能健康监护、人机交互等场景展现出重要应用价值。传统的深度学习算法，如卷积神经网络(CNNs)和循环神经网络(RNNs)，虽然取得了一定成效，但在捕捉复杂人体活动的时序动态和空间关系方面仍存在不足。为了解决这些局限性，文章利用Transformer模型在捕获全局特征方面的优势和胶囊网络在捕获局部特征方面的优势，提出了一种结合Transformer模型与胶囊网络的新型混合架构。在两个公共数据集(UCI-HAR和WISDM)上对模型性能进行了评估。最后，该模型在UCI-HAR数据集中的总体准确率为96.0%，在WISDM数据集中的总体准确率为96.5%。实验结果表明，基于Transformer和胶囊网络的融合模型比仅基于Transformer和仅基于胶囊网络的模型表现出更好的性能。而且，该算法的性能优于近期同类研究中的其他传统深度学习算法。

Abstract: In recent years, wearable sensor-based Human Activity Recognition (HAR) has demonstrated significant application value in intelligent health monitoring and human-computer interaction scenarios. Although traditional deep learning algorithms, such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), have achieved certain results, they still exhibit limitations in capturing the temporal dynamics and spatial relationships of complex human activities. To address these limitations, this paper proposes a novel hybrid architecture combining Transformer models and Capsule Networks, leveraging the Transformer’s advantages in capturing global features and Capsule Networks’ strengths in extracting local features. The model performance was evaluated on two public datasets (UCI-HAR and WISDM), achieving overall accuracies of 96.0% on the UCI-HAR dataset and 96.5% on the WISDM dataset. Experimental results demonstrate that the Transformer-Capsule hybrid model outperforms both standalone Transformer-based and Capsule Network-based models. Furthermore, the proposed algorithm exhibits superior performance compared to other traditional deep learning approaches in recent related studies.