The use of 3D marker-based motion analysis systems is considered the gold standard for tracking limb movements. However, these systems are expensive, limited to laboratory settings, and difficult to apply when studying paediatric populations. Therefore, this study investigated the validity of a markerless motion tracking software, DeepLabCut, in extracting kinematics from 2D sagittal plane videos of adults and toddlers during walking. Data were obtained from fifteen toddlers performing their first independent steps and sixteen healthy adults. Participants performed overground and treadmill walking at comfortable pace. Distinct models with either 25% or 75% of participants as input were used for DeepLabCut network training. For all participants, anatomical landmarks extracted through video analysis in DeepLabCut were used to calculate lower-limb gait kinematics, which were compared to kinematics computed using a 3D marker-based system (Vicon). The 25% models performed well for adult joint angles, but not for clinical parameters or toddlers. In both populations, the 75% models showed good (≥0.60) or excellent (≥0.75) intraclass correlation coefficient absolute agreement for most time-normalized joint angles and clinical parameters. This improvement was supported by increased Pearson’s correlation coefficients, decreased root mean squared errors, and increased R2 values from the 25% to the 75% models. More specifically, higher validity of DeepLabCut was found for adults compared to toddlers and treadmill compared to overground walking. Altogether, with sufficiently diverse input, DeepLabCut proved a valid tool to acquire gait kinematics in known scenarios, with the potential to study typical adult and toddler gait in more ecological and naturalistic environments.

DeepLabCut

Gait analysis

Markerless motion capture

Biomechanics

Deep learning

Toddlers

© 2025 The Author(s). Published by Elsevier Ltd.