Source-linked AI summary

Recovering Biomechanical Signals from Missing Keypoints Using Temporal Interpolation in Monocular Gait Analysis

Shubham Jariwala

arXiv:2609.09670v1cs.CVeess.IV

TL;DR

Monocular gait analysis can lose keypoints under occlusion, detection failures, or lightweight model constraints, while the sufficiency of simple temporal recovery remains unclear. This study tests first-order interpolation for missing ankle keypoints and finds that it restores knee-angle signals with high fidelity without learned reconstruction.

  • Problem

    Monocular gait analysis depends on keypoint reliability, but whether simple temporal filling can recover a biomechanically meaningful signal remains unclear.

  • Method

    The study compares baseline, persistent ankle-keypoint removal, and first-order exponential smoothing to recover the ankle before recomputing knee angles across a 527-frame walking sequence.

  • Results

    23.4° → 1.1° mean error after interpolation, with variance and smoothness returning to within a few percent of baseline.

  • Takeaways & Limitations

    A negligible-cost interpolation step can substantially mitigate occasional keypoint dropout and support low-complexity, real-time-compatible monocular gait analysis.

  • Takeaways & Limitations

    Generalizability is untested beyond one participant and one walking sequence, and performance under longer ankle-occlusion runs was not characterized.

Abstract

from arXiv · show

Monocular pose estimation enables low-cost gait analysis but is sensitive to missing keypoints caused by occlusion, detection errors, or efficiency-driven model reduction. While prior work on recovering missing joints focuses on complex learned models, the effectiveness of simple temporal methods remains underexplored. We evaluate knee-angle estimation under a missing-ankle-keypoint condition and test a first-order temporal interpolation scheme as a recovery mechanism. Across 527 frames of monocular walking video (428 with valid baseline detections), removing the ankle keypoint increased mean angular error to 23.4° +/- 46.7° and collapsed signal variance to near zero. Temporal interpolation reduced error to 1.1° +/- 6.7° and restored variance and smoothness to within a few percent of baseline. These results indicate that gait signals possess sufficient temporal redundancy for a simple, computationally trivial interpolation scheme to recover a critical missing joint, without resorting to learned reconstruction models. The findings support low-complexity, real-time-compatible designs for gait analysis in resource-constrained or occlusion-prone monocular settings.

1. Introduction

Monocular pose estimation offers low-cost gait analysis, but missing keypoints can undermine reliability. This study asks whether adjacent-frame temporal interpolation can recover a biomechanically meaningful knee-angle signal after ankle-keypoint loss.

  • Monocular pose estimation enables camera-only gait analysis without the expense and laboratory constraints of marker-based motion capture.
  • Keypoint reliability degrades under occlusion, motion blur, low-confidence detections, and lightweight models selected for real-time throughput.
  • Existing missing-pose recovery methods generally use learned priors, including Bayesian, geometric-constraint, temporal-context, three-dimensional, or deep-learning pipelines.
  • The study tests whether treating a missing keypoint as a temporal gap can recover a meaningful signal, using knee-angle estimation with ankle-keypoint loss.
  • The ankle is selected because it is commonly occluded in monocular views and directly contributes to knee-angle calculation.

2. Methods

The study computes knee angles from monocular pose detections and compares full, missing-ankle, and temporally interpolated conditions over a walking sequence. Evaluation uses valid baseline frames and measures angular error, variance, smoothness, and paired statistical differences.

  • Knee angle is computed at the knee from hip, knee, and ankle keypoints extracted by a YOLOv8-pose estimator.
  • The 527-frame sequence compares full baseline, ankle discarded in every frame, and first-order exponential smoothing for ankle recovery.
  • 99 of 527 frames lacked valid baseline keypoints, leaving 428 frames for recovery comparisons against the detected reference signal.
  • Evaluation measures mean angular error, signal variance, and smoothness as mean absolute frame-to-frame angle change.
  • The study uses one consenting author as the sole participant and collects no third-party participant data.

3. Results

Removing the ankle keypoint severely degrades knee-angle behavior, while temporal interpolation restores accuracy and temporal dynamics relative to the full baseline. Results are summarized across 428 valid frames and the three experimental conditions.

  • Table 1 reports angular error, variance, and smoothness for the full-baseline, missing-ankle, and interpolated conditions over 428 valid frames.
  • Ankle removal sharply increases error and collapses variance and smoothness toward zero, leaving the knee-angle trace effectively frozen.
  • Temporal interpolation returns variance and smoothness to within a few percent of baseline while substantially reducing point-wise error.
  • The missing-ankle and interpolated errors differ significantly under the Wilcoxon signed-rank test, with p ≈ 7.6 × 10^-59.

4. Discussion

The results support first-order temporal interpolation as a minimal recovery strategy for a missing ankle keypoint. They also suggest that occasional dropout can be mitigated without a learned reconstruction model.

  • First-order temporal interpolation is sufficient to recover a biomechanically meaningful knee-angle signal after loss of a load-bearing keypoint.
  • The findings suggest strong short-timescale gait redundancy, with an ankle position predicted from its immediately preceding position at typical video rates and walking cadences.
  • Robustness to occasional occlusion, low-confidence detections, or detector dropout does not necessarily require a learned reconstruction model.
  • A negligible-cost interpolation step can substantially mitigate this class of keypoint-dropout failure in monocular gait-analysis systems.

5. Limitations

The study's limitations concern generalizability, occlusion duration, and the absence of an independent motion-capture reference.

  • The study uses one participant and one walking sequence, leaving generalizability across gait patterns, camera angles, and occlusion durations untested.
  • The interpolation scheme was evaluated only for single-frame ankle loss; longer consecutive occlusions were not characterized and are expected to degrade performance.
  • Reference angles came from the same monocular pipeline rather than independent motion capture, so absolute errors indicate internal consistency rather than gold-standard accuracy.

6. Conclusion

Simple temporal interpolation recovered a knee-angle signal after loss of a load-bearing ankle keypoint, without a learned reconstruction model, supporting low-complexity monocular gait analysis.

  • 23.4° → 1.1° mean error was achieved when simple temporal interpolation recovered knee angle from a missing ankle keypoint.
  • The recovery operated without requiring a learned reconstruction model.
  • The result supports low-complexity, realtime-compatible gait analysis in occlusion-prone or resource-constrained monocular settings.

CRediT author statement

The CRediT statement assigns Shubham Jariwala responsibility across the study's conception, execution, analysis, and reporting.

  • Shubham Jariwala contributed to conceptualization, methodology, software, formal analysis, investigation, and data curation.
  • Shubham Jariwala contributed to writing the original draft and reviewing and editing the manuscript.
  • Shubham Jariwala also contributed visualization.
Loading 2609.09670v1…