Predicting walkway spatiotemporal parameters using a markerless, pixel-based machine learning approach

Abstract

BACKGROUND: Traditional gait analysis relies on motion capture systems and sensor mats, which are precise but costly and constrained to specialized environments. Markerless pose estimation combined with machine learning offers a low-cost alternative for gait analysis.

AIM: This study aimed to develop and validate a markerless method for predicting spatiotemporal gait parameters using video-based pose detection and machine learning, leveraging the GAITRite™ system as a reference.

METHODS: Ten healthy adults walked barefoot on a GAITRite™ mat while their movements were recorded. Videos were processed using the vailá Multimodal Toolbox, which integrates MediaPipe for pose estimation and various machine learning algorithms. Gait parameters such as step length, stride velocity, and support time were predicted using regression models trained on pixel-based data.

RESULTS: The results demonstrate a clear division in model performance based on the target variable. For metrics related to spatial characteristics, such as Step Length, Step Width, Stride Length, Stride Width, and Stride Velocity, the Gradient Boosting (GBR) delivered the best results, exhibiting lower error metrics and higher R² scores. Only for the Support Base data the best model differed, with the Random Forest (RFR) outperforming the others. This model also displayed lower error values and higher R² and EVS score.

INTERPRETATION: The results indicate the feasibility of markerless pixel-based gait analysis as a low-cost alternative to traditional methods, broadening the accessibility of precise gait assessment for both clinical and research applications.