Application of CNN-LSTM with Artificial Features and Attention Mechanism for Walking, Upstairs, and Downstairs Recognition
Year: 2026
Type: conference-paper
Abstract: Walking, going upstairs, and going downstairs are fundamental gait patterns in lower limb exoskeleton assistance systems, and accurate recognition of these activities is essential for adaptive control. However, distinguishing these patterns remains challenging due to the high similarity of acceleration signals, limited training samples, and class imbalance, especially between walking and going downstairs. To address these issues, this paper proposes a CNN-LSTM deep learning method that combines handcrafted features and an attention mechanism for gait activity recognition. The core focus of this work is on constructing a robust feature engineering pipeline, including signal preprocessing and multi-dimensional feature extraction, as well as optimizing the deep learning architecture for enhanced discriminative ability. Specifically, time-domain, frequency-domain, and kinematic parameters are extracted from inertial sensor data to form a feature vector with clear physical meaning. A CNN is used to capture local motion patterns, while an LSTM models the temporal evolution of gait cycles. A Squeeze-and-Excitation (SE) channel attention mechanism is further introduced to adaptively weight different feature channels, improving the model’s sensitivity to subtle differences between similar activities. Experimental results indicate that the proposed method attains an impressive 99.48% overall accuracy in classification, coupled with a $\mathbf{0. 9 9 3 0}$ macro-averaged F1-score. effectively reducing the confusion between similar gait activities. This work provides a reliable perceptual foundation for applications such as motion intention recognition and personalized assistance control in lower limb exoskeletons.
Institution: Changchun University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204914367