Markerless Gait Analysis and Circular Statistics: A Case Study in Parkinson’s Disease Classification

Markerless gait analysis enables Parkinson’s disease (PD) assessment directly from video, without markers or wearable sensors. Most pipelines describe joint-angle trajectories with standard summary statistics (means, standard deviations, . . . ). However, joint angles are directional quantities that lie on a non-linear manifold, and summarising them with statistics designed for data on a line may misrepresent their distribution. This paper investigates whether circular statistical descrip-tors provide a representation of angular gait signals that is at least as informative as the standard one for supervised PD classification. Using a public vision-based gait dataset and ViTPose for pose estimation, joint-angle trajectories are summarised with two matched feature families: a standard (linear) family and a circular one. The two families are com-pared under one identical supervised protocol that couples filter-based selection, recursive feature elimination, and four classifiers, assessed over repeated subject-level train–test splits. The goal is not to maximise clas-sification performance, but to test whether circular descriptors can match or improve upon the standard descriptors conventionally used for angular gait data. As a result, circular descriptor’s performance is comparable to, and in places better than, the standard ones, with the most visible differences in the more difficult trial and in the identification of PD subjects. When angular distributions are affected by pose-estimation errors, the adopted robust circular descriptors appear particularly useful. Estimates are obtained overrepeated splits, and differences should be read in light of run-to-run variability. Overall, results suggest circular statistics is a valuable, geometry-aware addition to markerless gait analysis.