A Dual-Stream CNN-LSTM Model for Human Activity Recognition Using Orthogonal Frequency Division Multiplexing Signals
Muneeb Ullah, Nan Zhao, Xiaodong Yang · 2025
Deep learning (DL) has been attracting major interest because of its remarkable performance in several fields. Its use in human activity recognition (HAR) with orthogonal frequency division multiplexing (OFDM) signals has not, however, been extensively investigated. Usually emphasizing spatial or temporal characteristics of signals separately, existing techniques neglect the interplay between these aspects. Using OFDM signals, this work presents a new CNN-LSTM based dual-stream architecture to identify human behaviors like walking, standing, and sitting. To capture both spatial and temporal relationships, the model makes use of the convolutional neural networks (CNN) and long short-term memory (LSTM) networks strengths. These models improve feature extraction, hence raising classification accuracy. With CNN and LSTM layers to record spatial and temporal attributes, each stream of the dual-stream model handles phase and amplitude information independently. By use of an outer product operation, the features from both streams interact to enhance the variety of obtained features and improve classification performance. Using OFDM signals, the experimental findings reveal that the proposed CNN-LSTM dual-stream structure much beats current state-of- the-art techniques, hence verifying its efficacy for HAR applications. The possibilities of DL-based techniques for non-invasive human activity categorization using wireless communication signals are emphasized in this paper.