⚡ electrical engineering

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

This paper demonstrates that simple first-order temporal interpolation can effectively recover missing ankle keypoints in monocular gait analysis, restoring knee-angle accuracy and signal variance to near-baseline levels without the need for complex learned reconstruction models.

Original authors: Shubham Jariwala

Published 2026-09-10✓ Author reviewed ⓘ

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shubham Jariwala

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Walking is a rhythm we rarely think about, a seamless flow of movement that our bodies perform without conscious effort. Yet, for scientists studying human motion, capturing this rhythm with a simple camera is a surprisingly difficult puzzle. For decades, researchers relied on expensive laboratories filled with reflective markers and specialized sensors to track how our joints move. Today, cheaper technology allows us to use a single video camera to estimate where our body parts are in space, a process known as pose estimation. This technology promises to bring high-quality movement analysis to clinics, gyms, and even homes. However, these camera-based systems are fragile. If a person's leg is briefly blocked from view, or if the software gets confused by motion blur, the system loses track of a specific point on the body, known as a keypoint. When a keypoint disappears, the calculation for how a joint bends often breaks down completely, turning a smooth, useful signal into static noise. The question facing researchers is whether this broken signal can be fixed with complex, heavy-duty computer models, or if a much simpler approach might work just as well.

In a recent study, a researcher at the Singapore University of Technology and Design tackled this problem by focusing on the knee, a joint that bends and straightens with every step. The researcher used a single video of a man walking and asked a computer to track his body using standard software. To test the system's weakness, the researcher deliberately removed the data point for the ankle in every single frame of the video, simulating a situation where the ankle is hidden from the camera. Without the ankle, the computer could no longer calculate the angle of the knee. The result was immediate and dramatic: the data for the knee angle stopped moving entirely. The signal flattened out, losing all the natural ups and downs of a walking step, and the error in the measurement jumped to an average of 23.4 degrees. The system had effectively frozen, unable to guess where the knee should be without the anchor point of the foot.

The researcher then tried a different approach, one that did not involve training a complex artificial intelligence to "guess" the missing data. Instead, they used a simple method called temporal interpolation. This technique looks at where the ankle was in the frame just before it disappeared and where it was in the frame just after, then fills in the gap with a smooth, calculated position based on those neighbors. It is a straightforward mathematical smoothing, similar to how one might estimate the position of a car between two known points on a road. When this simple fix was applied, the knee angle signal came back to life. The frozen line began to move again, regaining its natural shape and rhythm. The average error dropped from 23.4 degrees down to just 1.1 degrees, and the smoothness of the movement returned to nearly the same level as if the ankle had never been missing.

The study found that this simple fix worked because human walking has a strong, predictable rhythm over very short periods. The position of the ankle in the next split second is heavily influenced by where it was in the previous split second. This means that for a single missing frame, the body does not need a complex model to predict its future; it only needs to know its immediate past. The research suggests that for many real-world applications, such as monitoring gait in a busy clinic or a home setting, we do not need to build expensive, computationally heavy systems to handle occasional glitches. A basic, low-cost calculation can restore the signal with high fidelity. The findings indicate that the natural redundancy in how we walk is sufficient to allow simple software to recover from common errors like a lost keypoint, keeping the analysis running smoothly even when the camera view is imperfect.

Of course, this discovery comes with clear boundaries. The study was conducted on a single person walking in a single sequence, so it is not yet known if this simple fix works for every type of walk, every camera angle, or if it can handle a long period where the foot is hidden for many seconds at once. The researchers also noted that their "perfect" reference data came from the same camera system, meaning the results show how well the system can fix itself internally, rather than proving it matches a gold-standard laboratory measurement. Despite these limits, the work offers a reassuring path forward. It shows that in the quest to understand human movement with simple cameras, we may not always need the most powerful computers. Sometimes, the answer lies in the quiet, predictable rhythm of the step itself, which a simple calculation can hear even when the data is missing.

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

Problem Statement
Monocular human pose estimation offers a low-cost alternative to marker-based motion capture for gait analysis. However, its practical reliability is compromised by missing keypoints resulting from occlusion, motion blur, low-confidence detections, or the use of lightweight models optimized for real-time throughput. While existing literature often addresses missing joint recovery through complex learned models (e.g., Bayesian priors, temporal-context networks, or full 3D reconstruction pipelines), the efficacy of simple, computationally trivial temporal methods in recovering biomechanically meaningful signals remains underexplored. Specifically, it is unclear if treating a missing keypoint purely as a temporal gap filled by adjacent frames is sufficient to restore critical gait metrics.

Methodology
The study evaluates the recovery of knee-angle estimation under a simulated "missing-ankle" condition using a single participant (male, 1.70 m, 64 kg) recorded via monocular video.

Pipeline: A YOLOv8-pose estimator extracts 17 COCO-format keypoints. The knee angle ( $\theta$ ) is calculated as the angle at the knee joint between vectors formed by the hip, knee, and ankle.
Experimental Conditions: Three conditions were compared over a 527-frame sequence (428 frames retained after filtering for valid baseline detections):
1. Full Baseline: Knee angle computed from all detected keypoints.
2. Missing Ankle: The ankle keypoint is discarded for every frame, simulating hard occlusion or detector dropout.
3. Temporal Interpolation: The missing ankle position is recovered using a first-order exponential smoothing update ( $ankle'_t = 0.7 \cdot ankle'_{t-1} + 0.3 \cdot ankle_t$ ). The knee angle is then recomputed using this smoothed position.
Metrics: Performance was measured via mean angular error relative to the baseline, signal variance, and signal smoothness (mean absolute frame-to-frame change). Statistical significance was assessed using the Wilcoxon signed-rank test due to the right-skewed nature of per-frame error distributions.

Key Results
The removal of the ankle keypoint caused a catastrophic failure in signal quality. The mean angular error surged to 23.4° ± 46.7°, while signal variance collapsed to near zero (~0.0) and smoothness dropped to 0.01. This indicates that without the ankle vector, the calculated knee angle effectively freezes rather than merely becoming noisy.

In contrast, the first-order temporal interpolation scheme successfully restored the signal:

Error Reduction: Mean angular error decreased from 23.4° to 1.1° ± 6.7°.
Signal Restoration: Variance recovered to 2192.0 (compared to the baseline of 2177.8), and smoothness returned to 2.12 (baseline: 1.97).
Statistical Significance: The improvement was highly significant (Wilcoxon signed-rank, $p \approx 7.6 \times 10^{-59}$ ).

Key Contributions

Demonstration of Temporal Redundancy: The study provides empirical evidence that gait signals possess sufficient short-timescale redundancy such that a simple, first-order interpolation scheme can recover a critical, load-bearing joint (the ankle) with high fidelity.
Validation of Low-Complexity Recovery: The results challenge the necessity of complex learned reconstruction models for handling occasional keypoint dropout. A computationally trivial interpolation step was shown to restore both the accuracy and the underlying temporal dynamics of the gait cycle.
Real-World Baseline: The evaluation utilized a realistic detection rate (81.2% valid frames even without injected perturbations), measuring recovery performance against a real, imperfect reference signal rather than a synthetic ideal.

Significance and Claims
The paper concludes that simple temporal interpolation is sufficient to recover biomechanically meaningful knee-angle signals from missing keypoints in monocular settings. The author argues that this finding supports the design of low-complexity, real-time-compatible gait analysis systems for resource-constrained environments or occlusion-prone scenarios. The work suggests that robustness to keypoint dropout does not inherently require heavy deep learning reconstruction models, as a negligible-cost interpolation step can substantially mitigate this class of failure.

Limitations
The author explicitly notes that the study is limited to a single participant and a single walking sequence; thus, the generalizability of the specific error reduction magnitudes to other gait patterns, camera angles, or longer occlusion durations is untested. Furthermore, the "ground truth" is derived from the same monocular pipeline's detections rather than an independent motion-capture system, meaning the reported errors reflect internal consistency rather than absolute accuracy against a gold standard.

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

More like this