Prediction of human walking vertical load based on markerless human gait capture using multi-view cameras and LSTM network

• Published in: Engineering structures Vol. 334; p. 120228
• Main Authors: Liang, Huiqi, Xie, Wenbo, Lu, Yijing, Wei, Peizi, Zhang, Zhiqiang
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.07.2025
• Subjects: 3D keypoints reconstruction; Ground reaction force prediction; In-situ test; LSTM network; Multiview markerless gait capture; Pedestrian localization; In-situ test; 3D keypoints reconstruction; Pedestrian localization; Ground reaction force prediction; LSTM network; Multiview markerless gait capture
• ISSN: 0141-0296
• DOI: 10.1016/j.engstruct.2025.120228

The evolving trend toward large-span and flexible structures is driven by advancements in materials and construction methods. Recently, heightened attention has been directed towards structural vibration induced by human walking. Predicting walking loads for pedestrians, encompassing temporal and spatial aspects, is crucial. In-situ tests face challenges with intrusive methods, like labelling points or wearable sensors. This study proposes a markerless method for predicting human vertical walking force, utilizing multiple cameras, the OpenPose network for 2D key points, and the 3D Pictorial Structures (3DPS) method with the Skinned Multi-Person Linear Model (SMPL) human body model for optimized gait representation. Gait and walking force acquisition tests on 30 subjects employed a Long Short-Term Memory (LSTM) network for training and validation. In an outdoor space of 40 m x 15 m, in-situ tests evaluated accuracy in human walking force in time and frequency domains, and pedestrian localization on rigid ground. The proposed method exhibited a Root Mean Square Error (RMSE) of less than 0.05 g, a relative error below 5 %, and a first-order instantaneous synchronization coefficient exceeding 0.8 with inertial measurement unit (IMU) signals in vertical ground reaction force (GRF) prediction across all test cases. Pedestrian localization error between multi-view method and ultra-wideband (UWB) location system results was less than 15 cm, validating the feasibility and accuracy of markerless vertical spatial load prediction using human gait capture under multi-view conditions. •Markerless and Non-intrusive method to captures gait with multi-camera views.•Predicts pedestrian walking vertical force and location simultaneously.•Validates accuracy and applicability in a 40 m x 15 m in-situ test in outdoor space.•Enables long-term human-induced load monitoring in large-span structures.•Provides technic insights for gait capture in human-induced vibration in-situ tests.