A hybrid LSTM CNN model with efficient channel attention for enhanced human activity recognition using wearable sensors

Abstract Human activity recognition (HAR) is crucial for tracking human activity in various fields, including healthcare, context-aware computing, sports injury prevention, elder care, and home monitoring. Effective activity recognition supports patients with chronic diseases managed by healthcare professionals and encourages people to lead healthier lifestyles. In this paper, we have proposed a novel hybrid deep learning-based tempo-spatial architecture with adaptive cross-channel attention mechanisms, named TECA-HAR, which is particularly designed to improve human activity recognition performance. The proposed model effectively combines long short-term memory networks (LSTM) and convolutional neural networks (CNN) with the Efficient Channel Attention (ECA-Net) module to exploit the temporal and spatial aspects of time-series data. In contrast to traditional LSTM-CNN hybrids, our proposed hybrid model incorporates ECA-Net to assign channel-wise attention without reducing dimensions and improves feature selectivity without incurring computational overhead. In addition, a global average pooling (GAP) layer was employed instead of the traditional fully connected layer to minimize the model’s complexity and create a lightweight model while maintaining performance. We evaluated the proposed model using four public benchmark datasets: UCI-HAR, WISDM, PAMAP2, and DaphNet. The experimental results of the proposed model were more than satisfactory, with F1 scores of 96.74% in the UCI-HAR with six activities, 98.78% in the WISDM with six activities, 98.65% in the PAMAP2 with twelve activities, and 96.96% in the DaphNet with two activities, respectively. The proposed TECA-HAR model outperformed some baseline models (CNN, vLSTM, CNN-LSTM, BiLSTM, BiGRU, and DeepConvLSTM), and exhibited competitive performance concerning state-of-the-art (SOTA) approaches, reflecting its efficiency in various activity recognition tasks. Thus, the proposed lightweight framework has a high potential for use in real-time human activity monitoring systems, particularly in low-resource and embedded environments.