Multiscale Convolutional–Bidirectional LSTM Fusion of Spatiotemporal Attention for Human Activity Recognition

Abstract

Human activity recognition (HAR) is a core technology in fields such as smart healthcare and human–computer interaction, which aims to classify daily activities (e.g., walking and running) based on sensor data automatically. While existing approaches achieve high accuracy in controlled laboratory settings, they often perform poorly in real-world applications due to limited multiscale feature extraction and poor modeling of interchannel sensor correlations. These shortcomings lead to high sensitivity to noise and irrelevant time segments. An innovative model integrating multiscale convolution, bidirectional long short-term memory (BiLSTM), and a spatiotemporal attention mechanism is proposed to address these issues in this research. The model employs a multiscale parallel convolutional structure with filter sizes of 3, 5, 7, and 9 that enable it to capture both short-term local dependencies and long-term global patterns simultaneously. The introduction of a spatial–temporal dual attention mechanism dynamically focuses on key sensor channels and time segments, significantly improving the accuracy of feature selection. In addition, the Swish activation function is used to optimize the feature extraction process. Its smooth characteristics and self-gating mechanism avoid the gradient vanishing problem of the traditional ReLU effectively. The experimental results show that the model achieves an accuracy of 95.39% on the UCI–HAR public dataset and an excellent performance of 99.58% on the custom dataset.