Models that estimate posture, activity, or ergonomic risk from video, images, or skeleton sequences.
zheng-motionbert-stonemasonry-reba
Paper only
Huaqiao University · 2026
Cascaded markerless pipeline for an occupational trade with high-load manual tasks: YOLOv8n worker detection, knowledge-distilled RTMPose 2D keypoints, and MotionBERT spatiotemporal 3D reconstruction feeding automated REBA scoring (Measurement, 2026). Reports 28.3 mm mean per-joint position error under obstructed field conditions and 86.2 percent agreement with expert risk classifications for stonemasonry workers. No code release.
pose-estimationtransformermotionbertrebaconstructionmarkerlesscomputer-vision
InputRGB video of stonemasonry work, processed by YOLOv8n detection and distillation-optimized RTMPose 2D keypoints
OutputMotionBERT spatiotemporal 3D pose reconstruction and frame-wise REBA postural risk levels
DomainStonemasonry construction trade — automated postural risk assessment under cluttered field conditions
salehi-mobile-markerless-lifting
Paper only
Oregon State University / Texas A&M University · 2026
Adapts OpenCap, the open-source smartphone-based markerless motion capture platform, to occupational lifting by retraining its marker augmentation model on a large, diverse manual-lifting dataset (Appl. Ergon., 2026). Against reference motion capture, the task-specific model cut joint kinematic errors by about 37 percent versus the original OpenCap model (9.5 vs. 15.0 degrees RMSE), with gains concentrated in trunk and upper-body joints and robustness across lift heights, asymmetry angles, and camera setups.
markerlessmotion-captureopencapliftingsmartphonevalidationcomputer-vision
InputVideo from two or more mobile-device cameras, processed through the OpenCap markerless pipeline
OutputFull-body 3D joint kinematics during lifting, from a task-specific marker augmentation model
DomainManual lifting — field-viable 3D kinematics measurement for ergonomics applications
rs9000-ergonomics-reba-library
Code
Independent open-source project
Python library computing REBA scores directly from 3D pose keypoints (13+2 joints as root-relative X, Y, Z coordinates), released under the MIT license. A community open-source implementation of the REBA method (see the hignett-mcatamney-reba entry in this library) rather than a published model of its own: it bolts onto the output of any pose-estimation system to turn skeletons into postural risk scores, serving as a building block for automated occupational posture-risk pipelines.
rebaposture-assessmentopen-sourcepythonskeletonlibrary
Input3D pose keypoints (13+2 joints as X, Y, Z coordinates relative to the root joint): head, shoulders, elbows, wrists, hips, knees, ankles
OutputREBA scores computed from the joint configuration
DomainOccupational posture-risk pipelines — automated REBA scoring behind any pose-estimation system
lin-auto-method-selection-posture
Paper only
National Tsing Hua University · 2022
Computes joint angles from workplace video via OpenPose, then a decision tree picks the most appropriate observational method (REBA, RULA, or OWAS) from task repetitiveness and upper/lower-limb activity and scores MSD risk (Sci. Rep., 2022, open access). Tested on 15 operation videos spanning six occupational categories (assembly, maintenance, manual handling, cleaning, office work, driving), isolating the ~10% of frames that are high risk and cutting expert evaluation time by roughly 90%. No code release.
posture-assessmentopenposerebarulaowasdecision-treecomputer-vision
InputRGB video of work tasks, processed to OpenPose joint keypoints, no body-worn sensors
OutputAutomatically selected assessment method (REBA, RULA, or OWAS), its risk score, and flagged high-risk frames
DomainGeneral occupational tasks — automated observational posture assessment from workplace video
lee-scaffolding-3d-pose-risk
Paper only
Sungkyunkwan University / Gangneung-Wonju National University / KAIST · 2026
Vision-based framework estimating 3D joint angles from a single RGB camera, classifying construction-specific hazardous postures, and computing time-weighted cumulative REBA scores in real time for scaffolding work (ISARC 2026). Validation showed strong agreement with expert assessments, with stable detection of squatting postures and higher misclassification for overhead work due to depth ambiguity and self-occlusion. No code release.
pose-estimationmonocularconstructionscaffoldingrebareal-timecomputer-vision
InputMonocular RGB camera video of scaffolding workers, no wearable sensors
Output3D joint angles, construction-specific hazardous posture classes, and time-weighted cumulative REBA scores
DomainConstruction scaffolding work — continuous WMSD risk monitoring of high-risk postures
cruciata-lightweight-vit-posture
Paper only
University of Palermo / Polytechnic University of Marche / University of Perugia / University of Camerino · 2025
~5M-parameter vision transformer (SPECTRE-ViT) that classifies ergonomic risk for eight anatomical regions (neck, trunk, bilateral upper/lower arms, wrists) directly from raw RGB frames in a single multi-label pass (Sensors, 2025, open access). Trained on synchronized RGB video and full-body inertial motion capture with labels from RULA thresholds; reports F1 above 0.99 and AUC above 0.996 per region with real-time inference on edge devices. Code and data are available only on request to the authors.
posture-classificationvision-transformerdeep-learningrulawmsdmanufacturingedge-computing
InputRaw RGB video frames (224x224), no skeleton reconstruction or joint-angle estimation
OutputBinary ergonomic risk labels (acceptable vs. risky) for eight body regions per frame, from RULA-derived thresholds
DomainManual manufacturing — WMSD risk monitoring on assembly, handling, and quality-control tasks
chen-colearning-3d-pose-construction
Paper only
The University of Hong Kong · 2024
Lightweight monocular 3D human pose estimation model with a residual log-likelihood estimation head, trained with a co-learning scheme that draws 2D and 3D features from multi-dimension datasets simultaneously (Comput.-Aided Civ. Infrastruct. Eng., 2024). Combined with pose tracking, it recognizes workers' 3D postures in real time for on-site ergonomic risk assessment in construction. No code release.
pose-estimationmonocularconstructiondeep-learningreal-timecomputer-vision
InputMonocular RGB video of construction workers, no markers or wearable sensors
OutputReal-time 3D postures via pose tracking and ergonomic risk levels for monitored workers
DomainConstruction sites — continuous on-site ergonomic risk monitoring from a single camera
parsa-stcn-action-ergonomic-risk
Code
University of Washington / IIT Gandhinagar · 2019
IEEE RA-L 2019 method framing worker ergonomic-risk prediction as action segmentation of task video: per-frame VGG16 features drive an encoder-decoder temporal convolutional network that labels every frame with an action class tied to a safe, monitor, or high-risk tier, reaching 87-94% F1-overlap on videos over two minutes long. Introduced the UW-IOM dataset (20 participants manipulating boxes and rods; Mendeley Data, CC BY 4.0). TensorFlow code is on GitHub (no license stated).
activity-recognitionrisk-assessmentcomputer-visiondeep-learningtensorflowtemporal-convolutional-network
InputRGB video frames; VGG16 spatial features fed to a temporal convolutional network
OutputFrame-wise action labels mapped to safe / monitor / high-risk ergonomic tiers
DomainIndoor object manipulation (warehouse-style pick-and-place) — ergonomic risk prediction
parsa-mtl-era-activity-reba
Code
University of Washington · 2021
Multi-task graph convolutional network (WACV 2021) that jointly segments worker activities and regresses frame-wise REBA ergonomic-risk scores from 3D skeleton sequences: a shared GCN backbone feeds an encoder-decoder TCN segmentation head and an LSTM risk head. On the UW-IOM dataset the best variant reached 92% F1-overlap for segmentation with a REBA-score MSE of 0.61, versus 1.68 for the single-task baseline. PyTorch code is on GitHub (no license stated).
activity-recognitionrebarisk-assessmentgraph-neural-networkdeep-learningpytorchskeleton
InputSequences of 15-joint 3D skeleton keypoints from long task videos
OutputFrame-wise activity segmentation labels and continuous REBA ergonomic-risk scores
DomainIndustrial object-manipulation work — activity recognition and postural risk scoring
mehrizi-multiview-3d-lifting-pose
Paper only
Rutgers University / North Carolina State University / UNC Charlotte · 2019
Two-stage deep network estimating full-body 3D pose during lifting from two ordinary camera views - an hourglass 2D module per view feeding a multi-view 3D pose generator (J. Biomech., 2019; the arXiv preprint is the FG 2018 conference version). Validated against synchronized marker-based motion capture on 12 adults performing nine lifting variations (three heights, three asymmetry angles): mean 3D joint error 14.7 +/- 3.0 mm, with larger errors at 60-degree asymmetry and shoulder height. The lifting dataset and code were not released.
pose-estimationmarkerlessliftingdeep-learningmulti-viewcomputer-vision
InputRGB video frames from two synchronized camera views at 90 and 135 degrees, no markers or depth sensors
OutputPer-frame 3D coordinates of 14 body joints
DomainManual lifting — markerless full-body 3D pose for biomechanical analysis
greene-cv-rnle-trunk-kinematics
Paper only
University of Wisconsin-Madison / NIOSH · 2021
Extends the Revised NIOSH Lifting Equation with trunk kinematics extracted from ordinary workplace video (HFES Proceedings, 2021): reanalyzing videos from NIOSH's prospective 78-worker cohort, average trunk speed and acceleration correlated with low-back-pain outcomes, and adding them to the Lifting Index significantly improved risk prediction (p = 0.003). The underlying trunk-angle extraction method (linked as the record; Human Factors, 2022, free full text) was validated against motion capture on 216 lifting trials (mean absolute difference 14.7 degrees, r-squared = 0.80).
computer-visionliftingniosh-lifting-equationtrunk-kinematicsrisk-assessmentmarkerless
InputConventional 2D workplace lifting video processed with a computationally efficient computer-vision method
OutputAverage trunk angular speed and acceleration, combined with the RNLE Lifting Index in a logistic low-back-pain risk model
DomainManual lifting — video-based low-back-pain risk prediction
cp3d-construction-pose-risk
Code
University of Alberta · 2024
Construction-specific 3D pose resource and model (Automation in Construction, 2024, open access): ~420k Vicon-captured 3D poses of 7 actors across 14 common construction activities, on a purpose-built skeletal model that supplies every joint angle REBA and RULA require, plus a PyTorch multi-person pipeline scoring worker ergonomic risk from RGB images. Models trained with CP3D outperform those trained without it. The GitHub repo carries a REBA demo (no license stated); the dataset itself is requested via a link in the README.
pose-estimationconstructionrebarularisk-assessmentdeep-learningpytorchmocap
InputRGB images of construction workers with bounding-box and root-depth information
Output3D joint coordinates on a biomechanical skeleton covering all REBA/RULA joint angles, plus REBA and RULA risk scores
DomainConstruction work — 3D pose estimation and ergonomic risk assessment
agostinelli-cv-ergonomic-assessment
Paper only
Universita Politecnica delle Marche / University of Macerata · 2024
Markerless ergonomic-assessment platform combining tf-pose-estimation and MediaPipe to track joint angles from work-cycle video and compute REBA, RULA, and OCRA — validated in situ on three real manufacturing production lines with 12 operators against expert ergonomists (Scientific Reports, 2024, open access). Risk-level agreement reached 40-80% depending on line and method; the authors position it for continuous monitoring alongside, not instead of, expert assessment. The linked GitHub repo holds the raw validation data; the tool's source code is not released.
computer-visionrebarulaocramarkerlessrisk-assessmentmanufacturingfield-validation
InputRGB video of work cycles from commodity action cameras behind and lateral to the workstation
OutputJoint-position and inter-limb-angle time series with REBA, RULA, and OCRA scores over the work cycle
DomainManufacturing production lines — observational postural risk scoring from video
yen-radwin-mvta-video-task-analysis
Software
University of Wisconsin-Madison · 1995
Video-based data acquisition and interactive data-extraction method (IEEE Trans. Biomed. Eng., 1995): analog biomechanical signals are recorded in synchrony with task video, and analysts mark event breakpoints while the video plays at any speed or direction to obtain elemental times, event frequencies, duty cycles and postures. Implemented as the MVTA software, distributed commercially by NexGen Ergonomics, which supplied the duty-cycle measurements behind the frequency- and hand-speed-based ACGIH HAL equations.
video-analysistime-studyduty-cyclehand-activity-levelupper-extremity
InputTask video with analyst-marked event breakpoints; optional synchronized analog channels (EMG, goniometers)
OutputElemental times, event frequency, duty cycle, postural analysis
DomainRepetitive manual work — time study and exposure assessment from task video
wang-2d-video-lifting-monitor
Software
University of Wisconsin-Madison / NIOSH · 2019
Single-camera computer-vision method that detects lifting instances with a motion "ghosting" effect and predicts hand and feet locations to extract the spatial and temporal factors of the revised NIOSH lifting equation (Ergonomics, 2019); in the lab its recommended-weight-limit estimates were within 0.2 kg of 3D motion capture. Builds on bounding-box lifting-posture classification (Greene et al., Human Factors, 2019); patented as US 11,450,148 (WARF) and commercialized by KineVid LLC.
computer-visionliftingniosh-lifting-equationmarkerlessdeep-learning
InputSingle sagittal-plane video camera, no sensors on the worker
OutputLift detection, hand and feet locations, and the horizontal/vertical RNLE factors for the recommended weight limit
DomainManual lifting — Revised NIOSH Lifting Equation risk assessment
chen-video-hand-activity-level
Software
University of Wisconsin-Madison · 2013
Marker-less digital video processing that tracks hand motion in ordinary workplace video and quantifies repetitive hand activity automatically (Human Factors, 2013), later extended to fully automated elemental-time and duty-cycle measurement (Akkas et al., Ergonomics, 2016). Together with the hand speed-duty cycle equation it forms Radwin's video HAL pipeline, patented as US 9,566,004 and commercialized by the UW-Madison spin-off KineVid LLC.
computer-visionhand-activity-levelmarkerlessduty-cycleupper-extremity
InputConventional 2D workplace video; marker-less hand tracking
OutputHand speed, duty cycle and exertion frequency, converted to an ACGIH hand activity level
DomainRepetitive hand-intensive work (upper-extremity exposure assessment)
wang-roofing-pose-estimation
Paper only
NIOSH / University of Delaware · 2022
NIOSH-University of Delaware method for estimating 3D worker pose directly from video of residential roofing tasks (Comput. Methods Biomech. Biomed. Eng.: Imaging & Visualization, 2022), a step toward markerless ergonomic exposure assessment on job sites where marker-based motion capture is impractical. Public as a paper and NIOSH record; no code release.
pose-estimationmarkerlessroofingconstructiondeep-learning
InputRGB video of roofing work
Output3D human pose (joint positions)
DomainResidential roofing (construction)
NIOSH / University of Delaware · 2023
Deep-learning model that estimates whole-body joint kinematics directly from multi-view images (Computer Vision and Image Understanding, 2023), developed in a NIOSH-University of Delaware collaboration for markerless motion analysis of occupational tasks such as roofing. Open-access preprint on arXiv; PyTorch code released on GitHub under MIT.
pose-estimationmarkerlesskinematicsroofingdeep-learningpytorch
InputMulti-view RGB images
OutputWhole-body joint kinematics
DomainOccupational tasks (residential roofing)
