Source-linked AI summary
Your Turn: At Home Turning Angle Estimation for Parkinson's Disease Severity Assessment
Qiushuo Cheng, Catherine Morgan, Arindam Sikdar, Alessandro Masullo, Alan Whone, Majid Mirmehdi
TL;DR
Clinical ratings provide brief snapshots despite fluctuating PD symptoms, motivating more frequent, objective measures of gait. The paper estimates turning angles from monocular-video poses and reports a significant group difference in maximum angular velocity, with PD subjects turning more slowly.
Problem
Sensitive, frequent, objective biomarkers are scarce for measuring slow PD progression alongside the current clinical rating scale.
Source: §1 · p. 1Method
Results
Cross-sectional analysis found a statistically significant group difference in maximum angular velocity: PD subjects turned much slower and showed subtle turning-angle reductions.
Source: §7 · p. 10Takeaways & Limitations
The work establishes a baseline for future passive, video-based analysis of turning in challenging free-living environments and potential future longitudinal in-home monitoring of PD.
Source: §8 · p. 10The authors state that current accuracy may not yet allow real-world application; the study included 24 people, 12 with PD.
Source: §8 · p. 10
Abstract
from arXiv · showhide
People with Parkinson's Disease (PD) often experience progressively worsening gait, including changes in how they turn around, as the disease progresses. Existing clinical rating tools are not capable of capturing hour-by-hour variations of PD symptoms, as they are confined to brief assessments within clinic settings. Measuring gait turning angles continuously and passively is a component step towards using gait characteristics as sensitive indicators of disease progression in PD. This paper presents a deep learning-based approach to automatically quantify turning angles by extracting 3D skeletons from videos and calculating the rotation of hip and knee joints. We utilise state-of-the-art human pose estimation models, Fastpose and Strided Transformer, on a total of 1386 turning video clips from 24 subjects (12 people with PD and 12 healthy control volunteers), trimmed from a PD dataset of unscripted free-living videos in a home-like setting (Turn-REMAP). We also curate a turning video dataset, Turn-H3.6M, from the public Human3.6M human pose benchmark with 3D ground truth, to further validate our method. Previous gait research has primarily taken place in clinics or laboratories evaluating scripted gait outcomes, but this work focuses on free-living home settings where complexities exist, such as baggy clothing and poor lighting. Due to difficulties in obtaining accurate ground truth data in a free-living setting, we quantise the angle into the nearest bin $45^\circ$ based on the manual labelling of expert clinicians. Our method achieves a turning calculation accuracy of 41.6%, a Mean Absolute Error (MAE) of 34.7{\deg}, and a weighted precision WPrec of 68.3% for Turn-REMAP. This is the first work to explore the use of single monocular camera data to quantify turns by PD patients in a home setting.
1. Introduction
Clinical assessments provide only brief snapshots of fluctuating PD symptoms, motivating sensitive, frequent, objective gait measures. The introduction presents a monocular-video pipeline and datasets for estimating turning angles in home and controlled settings.
1. Introduction
- Clinic-based MDS-UPDRS assessments provide brief snapshots even though PD symptoms fluctuate hourly, while sensitive, frequent, objective biomarkers of slow disease progression remain scarce.The scale also includes subjective questionnaires and clinician ratings of scripted activities.Source: §1 · p. 1
- Turning abnormalities are common in PD, and turning angles may help characterize medication status, freezing of gait, fall risk, and other turn measurements.Turning parameters may also be useful in clinical trials, as turning gait parameters can be more sensitive to change than straight-ahead gait outcomes in unmedicated early-stage PD.Source: §1 · p. 2
- The proposed pipeline extracts 3D joint coordinates from monocular RGB video and estimates turn angle from the orientation of paired hip and knee joints.The authors use the frontal plane and quantise angles into the nearest 45° bin because clinicians cannot reliably estimate precise degrees by eye.Source: §1 · p. 2
- The study also curates Turn-H3.6M from Human3.6M for comparative evaluation, with 3D data enabling turning-speed calculation.The curated dataset comprises 619 turning clips recorded under controlled settings.Source: §1 · p. 2
- The authors describe this as the first computer-vision approach to measure PD turning angles in free-living video without motion-capture reference.They position the work as the first to estimate turning angles from natural free-living video of people with PD.Source: §1 · p. 2
- Turn-REMAP provides free-living home turning videos from people with PD and healthy controls, with turning-angle labels discretised by expert clinicians.The dataset comprises turning clips from unscripted, passively collected home-monitoring videos.Source: §1 · p. 2
2. Literature review
Prior work offers precise sensor-based measures in restricted settings, while video-based gait analysis faces accuracy and generalisation challenges. The review identifies a need to validate absolute turning-angle measures in representative, free-living PD videos.
2. Literature review
- Wearable algorithms can reach sub-degree accuracy against motion capture or human raters in scripted laboratory turns, but their burden and poor lab-to-home translation constrain free-living use.Wearables require charging, reconfiguration, and manual data transfer, and imperfect use in uncontrolled settings can hinder generalisation.Source: §2 · p. 3
- Wearables provide kinematics for only a few body parts, and placement changes can produce inconsistent turn angles; video is passive but joint-angle accuracy and reproducibility remain concerns outside laboratories.The review calls for video datasets representative of different patients and settings, including homes.Source: §2 · p. 3
- Prior video-based PD gait studies use pose-derived features to infer MDS-UPDRS scores, while the review calls for quantitative validation of measured gait parameters.Examples include handcrafted features, classifiers, and deep-learning models applied to clinical annotations.Source: §2 · p. 3
- Earlier pose-based angular gait work focused on sagittal or coronal joints, while turning studies addressed turn detection or other gait metrics rather than turning angle.The review argues for measuring absolute gait parameters in natural, non-hospital settings.Source: §2 · p. 3
3. Datasets
Turn-REMAP provides turning clips from PD and control participants recorded in a home environment, with clinician annotations; Turn-H3.6M adds turning clips with motion-capture ground truth from a controlled setting.
Turn-REMAP
- Turn-REMAP comprises turning actions from 24 people living in a sensor-embedded home: 12 people with PD paired with 12 healthy controls.The videos include loosely scripted and spontaneous activities, as well as clinical assessments.Source: §3 · pp. 3–4
Angle labels
- Turn-REMAP angles are grouped into 45° bins, such as assigning angles from 67.5° to 112.5° to the 90° bin.The Turn-H3.6M distribution uses ground-truth angles calculated from changes in body orientation based on hip joints projected onto the ground plane.Source: Figure 2 · p. 4, §3 · p. 4
Turn-REMAP
- Clinicians annotated turn angle in 45° increments and turn duration, defining an episode from pelvis-rotation onset through movement completion.The two raters had Cohen’s kappa of 0.96 for angle annotations.Source: §3 · p. 4
Turn-H3.6M
- Turn-H3.6M provides 3D human-pose ground truth, with hip orientation used as the turning-angle reference and enabling turning-speed calculations.The source benchmark contains RGB and 3D data from professional actors performing activities in a customised laboratory environment.Source: §3 · p. 4
- Turn-H3.6M contains 619 turning clips selected from Human3.6M using consecutive hip rotation of at least 45° as the turning threshold.Its 45,199 frames were recorded at 50 fps; clip angles range from 45.2° to 234.7°.Source: §3 · p. 4
4. Methodology
The method estimates 3D skeletons from monocular video, then calculates turning angle from projected body-joint vectors. It uses temporal context to reconstruct pose and evaluates estimates against clinician annotations.
3D human joints estimation
- The pipeline detects 2D joints in each video frame and lifts the skeleton sequence into 3D using a two-stage pose-estimation framework.FastPose detects 2D keypoints; the Strided Transformer reconstructs 3D joints from temporal 2D skeletons.Source: §4 · pp. 5–6
- The Strided Transformer uses temporal smoothness constraints and long-range sequence information to reconstruct 3D joints.Its central-frame predictions draw on surrounding frames; the described context window is 27 frames.Source: §4 · p. 6
Turning angle estimation
- Turning angle is estimated from hip and knee joints on the frontal plane, projected onto the assumed flat ground plane.The method calculates changes between consecutive-frame vectors and averages the hip- and knee-based angles.Source: §4 · p. 6
- Angular speed is calculated from the turn angle and video duration, while frame-by-frame estimates can also show velocity changes within a turn.For overall angle, the frame-by-frame calculation is mathematically equivalent to using only the first and last frame vectors.Source: §4 · p. 6
Evaluation
- The angle-estimation algorithm can be paired with any 3D pose model that produces skeletons, and hip-plus-knee vectors perform best among the tested joint combinations.Evaluation compares estimates with clinician annotations using accuracy, weighted precision, and mean absolute error.Source: §4 · p. 6
5. Experiments
The pipeline was evaluated against clinician labels on 1,386 Turn-REMAP turns and against 3D ground truth on Turn-H3.6M. Performance varied with setting and activity, while the cross-sectional analysis found a significant group difference in maximum angular velocity but not turning angle.
Results on Turn-REMAP / Results on Turn-H3.6M
- 41.6% accuracy, 34.7° MAE, and 68.3% WPrec were achieved across 1,386 Turn-REMAP videos; Turn-H3.6M yielded 73.5% accuracy, 18.5° MAE, 15.5°/s turning-speed MAE, and 86.2% WPrec.Turn-REMAP predictions were compared with clinician annotations; Turn-H3.6M provided 3D ground truth.Source: §5 · pp. 6–7, Table 4 · p. 7, Table 5 · p. 8
Results on Turn-REMAP
- Clinical-assessment turns performed worse than loosely scripted and free-living turns, with self-occlusion cited as a major factor; 40 of 49 clinical turns were scripted 180° turns in a narrow hallway.Across scenarios, accuracy ranged from 26.5% to 44.0%, MAE from 33.4° to 59.2°, and WPrec from 66.2% to 79.1%.Source: §5 · pp. 6–7
- Turn-REMAP location-level accuracy ranged from 35.9% to 42.9% and averaged 38.6%, with occluded dining-room and hallway settings associated with higher MAE and lower accuracy.The location-level MAE averaged 33.9° and WPrec averaged 71.4%.Source: §5 · p. 7
Results on Turn-H3.6M
- Turn-H3.6M angle-prediction accuracy averaged 71.6% across angle bins and 74.1% across subjects; held-out subjects S9 and S11 fell within the range observed for training subjects.The angle-bin and subject-group results are reported separately; the paper describes the held-out performance as suggesting potential generalisation to unseen data.Source: §5 · p. 7, Table 5 · p. 8
- Across different actions, Turn-H3.6M predictions averaged 75.1% accuracy, 17.7° MAE, 14.5°/s turning-speed MAE, and 87.9% WPrec.The paper notes that movement dynamics, including speed and motion pattern, may contribute to performance differences across activities.Source: §5 · p. 7
Comparison between performance on Turn-REMAP and Turn-H3.6M
- Turn-H3.6M predictions were closer to bin angles and less variable than Turn-REMAP predictions; 89.5% of Turn-H3.6M errors were below 40°.Turn-REMAP standard deviations across the examined bins were 35.3°, 37.6°, 46.4°, and 70.0°, versus 22.2°, 20.5°, 16.8°, and 21.5° for Turn-H3.6M.Source: §5 · p. 7, Figure 5 · p. 8
Cross-sectional analysis of free-living turning characteristics
- The PD group’s mean turning angle was 92.65° versus 103.75° for controls, a nonsignificant difference (p=0.0998), while maximum angular velocity differed significantly (p=0.0339).Mean maximum angular velocity was 127.86°/s for PD and 160.19°/s for controls.Source: §5 · pp. 7–8, Table 6 · p. 8
6. Ablations
The ablations examine the choice of 2D keypoint detector and the joints used to calculate turning angle. FastPose offered the highest accuracy and lower computational cost among tested detectors, while joint combinations traded off accuracy and MAE.
Ablation study
- The ablation identifies accurate 2D keypoint detection as an important contributor to overall method accuracy.The authors frame both keypoint detection and body-part selection as design choices affecting turning-angle estimation.Source: §6 · p. 8
The effect of different 2D keypoints
- FastPose had the highest turning-angle accuracy among the tested 2D keypoint detectors and significantly reduced keypoint-detection computational costs.The three detectors’ reported MAE values were 18.4° and 18.5°, with no significant variation stated.Source: §6 · pp. 8–9, Table 7 · p. 9
The effect of using different joints
- Using both hip and knee joints gave the best average predicted-angle accuracy, while averaging hip, knee, and shoulder joints gave the lowest turning-angle MAE.The authors note that shoulder joints may be less reliable for initiating turns in PD because of greater frontal-plane inclination during weight shifts.Source: §6 · p. 9, Table 8 · p. 9
7. Discussion
The discussion reports group-level turning differences in Parkinson’s disease while identifying occlusion and cross-dataset generalization as important limits on angle estimation. It also frames the current accuracy as a baseline rather than a diagnostic tool.
7. Discussion
- Turn-REMAP accuracy was not yet robust enough for clinical diagnosis, with depth-reconstruction ambiguity and generalization to unseen in-the-wild poses remaining substantial challenges.The authors describe the ill-posed nature of recovering depth from 2D images as one factor affecting angle calculations.Source: §7 · p. 9
- Turn-REMAP performance was better for free-living and loosely scripted turns than clinical-assessment turns; 40 of 49 clinical turns were scripted 180° turns in a narrow hallway.Greater occlusion in locations such as dining rooms and halls corresponded to lower accuracy and higher MAE; PD and control prediction performance showed no large difference.Source: §7 · p. 9
- Generalizing pose models pretrained on lab-based Turn-H3.6M to in-the-wild Turn-REMAP is challenging amid differences in position distributions, cameras, body shapes, and movements.The discussion calls for domain adaptation and cross-dataset validation.Source: §7 · p. 9
- PD subjects showed statistically significant differences in maximum angular velocity, slower turns, and subtle angle reductions, despite pose models being pretrained on subjects without PD.The observed gait impairment was consistent with findings in other PD cohorts.Source: §7 · p. 10
- 73.5% accuracy on Turn-H3.6M; pose-estimation models optimized with MPJPE do not sufficiently address temporal smoothness or accurate turning-angle estimation.The authors identify downstream turning analysis using extracted deep-learning features as a direction for further work.Source: §7 · p. 10
8. Conclusion and future work
The paper introduces two datasets and a 3D-pose-based pipeline for estimating turning angles, with Turn-REMAP focused on clinician-annotated free-living turns. The authors present the work as a proof of concept and baseline, while noting that current accuracy may not support real-world application.
8. Conclusion and future work
- Turn-REMAP provides clinician-annotated, quantised turning angles for PD and control participants in free-living settings, while Turn-H3.6M adapts Human3.6M for turning analysis.The framework reconstructs 3D joints from RGB video and calculates turning angle from joint rotation.Source: §8 · p. 10
- With 24 participants, including 12 people with PD, the work is a proof of concept and establishes a baseline, but current model accuracy may not yet allow real-world application.The authors describe automatic free-living angle measurement as foundational for future longitudinal, in-home PD monitoring.Source: §8 · p. 10
- The datasets contain trimmed turning clips; proposed extensions include untrimmed video, additional metrics such as turning speed, and alternative skeleton models.The authors suggest these metrics could support classification or inference of clinical scores and medication status in free-living recordings.Source: §8 · p. 10