Answers are grounded in this paper and cite their sources.
In this article, we present our collected dataset with hardware-synchronized motion capture and depth sensors of a freely moving subject, our estimation of human pose, stride length, step length, and step length classification using deep learning, classical machine learning, and established algorithms. Our results on the 157,825-frame dataset show that pose estimation can be done with up to 85.91% of correct keypoints and as low as 8.86 cm mean per key point error with 2–3 cm of that error attributed to camera imprecision due to the 2–4 m distance. The stride estimation achieves up to 99.58% step percent and 100.39% distance ratio. The center of mass and foot distance based step length estimations show very similar step counts with 1380 and 1396, respectively, but differ in the total distance traveled at 58 m. The step length classification works well with an equal class recognition spread at an overall 80% accuracy. The core contributions of this work are the developed pose estimation models and the evaluation of all machine learning models and algorithms, the pipeline and evaluation scheme for every step from depth sequences to stride and step length estimation in a scenario of a freely moving subject, and the in-depth analysis tying the signals, algorithms, models, and gait phases together and highlighting the importance of different joint sets for step length classification.
No takes yet. Share an insight, caveat, or question.
Hartmann et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: