Prediction of human-induced structural vibration based on multi-view and markerless human gait capture and BiLSTM network

Abstract

Human-induced structural vibration assessment is important to evaluate the vibration serviceability of a structure. Existing test methods are mostly invasive, using wearable sensors for the direct measurement of pedestrian vertical loads in vibration test. However, invasive acquisition of a nonspecific tester is impractical for the purpose of monitoring and prediction of structural vibration. Moreover, there is a lack of practical tests considering the pedestrian gait evolution and trajectory, essential for accurate vibration prediction. For this purpose, this paper presents a non-invasive method for reconstructing pedestrian gait based on multiple camera views. Combining the Skinned Multi-Person Linear Model (SMPL) human body model with a Bidirectional Long Short-Term Memory (BiLSTM) network, a mapping network for “gait-walking force” was trained. The method achieves non-invasive and non-contact prediction of human walking loads. Experimental validation on a 7.5 m × 5 m test platform confirmed the method’s precision in predicting both walking loads and structural vibrations.