SMAGNet: Scaled Mask Attention Guided Network for Vision-based Gait Analysis in Multi-person Environments

Clinical gait analysis plays a key role in diagnosing and managing neurodegenerative diseases such as Parkinson’s disease. In recent years, vision-based gait analysis methods have emerged as promising non-invasive approaches to quantify gait characteristics. However, most methods assume single-person situations, but multi-person situations are more common in real-world medical settings. In this paper, we propose a novel mask-guided attention model called a Scaled Mask Guided Attention Network (SMAGNet), which exploits a target person's detection result to address multi-person issues. SMAGNet utilizes a detection box as a mask label to predict attention maps that highlight patients’ gait features and progressively refines the maps for accurate analysis. Experimental results show that the mean absolute percentage error (MAPE) was improved by up to 20% for the target spatio-temporal gait variable compared to the baseline 3D CNN (Convolutional Neural Networks). Moreover, we achieved significantly better performance compared to other methods, including a recent state-of-the-art gait recognition model named GaitBase. These results showcase SMAGNet’s effectiveness in multi-person gait analysis and its potential for real-world clinical use.