Measurement Error of Markerless Motion Capture Systems Applied to Tracking Movements in Human–Object Interaction Tasks: A Systematic Review with Best Evidence Synthesis

Abstract

This systematic review focused on the validity of markerless motion capture (MMC) systems used for human movement assessment during tasks that involve physical interaction with objects. Five electronic databases were searched until May 2025. Eligible studies (i) assessed the validity of an MMC system, (ii) required human participants to perform tasks that involved physical interaction with objects (e.g., lifts, carrying, gait with loads), (iii) employed a marker-based reference system, and (iv) reported at least one kinematic metric. Risk of bias was assessed using the SURE checklist. A best-evidence synthesis was conducted to classify the level of evidence across included studies. Fifteen studies met eligibility (median = 21 participants per study). In general, MMC systems presented good performance in capturing the waveforms related to movement (i.e., high associations with reference systems), but its level of precision (i.e., the magnitude of differences to the reference systems) still requires improvement regarding tasks involving human–object interactions. Most tasks analyzed were lifts, gait with load, squatting and reaching/manipulation, and technical gestures. There was strong evidence for the validity of MMC for implementation during lifting tasks. In summary, markerless motion capture (MMC) systems exhibit promising evidence of validity for some human–object interaction tasks, that is, especially when lifting as strong evidence was observed across studies on this type of task. In contrast, some evidence for tasks including gait under load, squatting, reaching, or touchscreen interaction is limited, moderate, or conflicting. Notwithstanding these limitations, most studies were observed to have moderate- to high-quality methodology. Additional research is required to optimize protocols to study the measurement error aspects of MMC under human–object interaction in real-world environments.

1. Introduction

Markerless motion capture (MMC) refers to the process of tracking human movement directly from video or depth images using computer vision and machine learning without reflective markers attached to the body of evaluated subjects [1,2]. Researchers and practitioners in movement science are increasingly interested in these systems due to their less invasive nature, improved portability, and often lower cost as compared with marker-based systems [3]. There are several validation studies and even recent reviews and meta-analyses reporting adequate or excellent accuracy and reliability for spatiotemporal movement measures (e.g., gait velocity and step length) for MMC systems [4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19]. While there is a knowledge base from such previous review studies, most of them were conducted with simple, formally controlled experiments in mind, and none have solely examined movements under conditions of some form of human–object interaction (e.g., lifting, carrying, or completing overall manipulation with tools), that may represent the physical actions requested in real-world tasks [20]. Thus, collated evidence specifically about the performance of MMC during human–object interactions tasks is currently lacking.

Measuring human movement features in the tasks requiring interactions with objects can increase the usefulness of MMC systems for a range of distinct fields (e.g., [21,22,23]). In ergometrics, previous work evaluated the utility of MMC for lifting and handling tasks in adding in-depth detail to distinguish occupational risks and posture assessment [24,25,26]. Markerless systems have been also implemented for assessing functional tasks within clinical rehabilitation contexts, such as reaching and upper-limb manipulation [27]. Of note, it is possible to observe that MMC has been now adopted for monitoring patients outside laboratory environments [6,28]. In sports and performance, MMC is being used to conduct ecologically valid assessments of skill execution, ranging from gait [9], skiing [10] and tactical movements (e.g., box jumps, single-leg landings, squat, or pivoting under heavy load) [11,13]. Collectively, these applications demonstrate the promise of MMC for escaping the control of laboratory settings and achieving real-world clinical relevance. However, measuring human movement kinematics with MMC while the subjects interact with objects could still involve problems in identifying the regions of interest similar to those of traditional marker-based kinematic systems. These include, for example, the occlusion effect derived from the potential overlapping of the object relative to the body in the image sequences, in addition to potential variations in lighting in the environment, which can also impact the MMC system functioning [20,29,30].

Despite increased validation of MMC systems, reviews to date have mainly summarized simple tasks, such as unladen gait or posture in laboratory environments [4,5,6]. There is limited research evaluating their use in tasks involving human–object interaction given the relevance of these situations in ergonomics, clinical, and sport-related settings, and this gap highlights the need for a systematic synthesis of the evidence on this topic. Therefore, the purpose of this systematic review was to examine the validity of markerless motion capture systems for tasks involving human–object interaction.

2. Materials and Methods

The protocol for this review was carried out in accordance with items suggested by The Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA 2020) statement [31], and was registered in the OSF—Open Science Framework under ID #meugz (https://osf.io/9zx6c; (accessed on 15 September 2025), DOI: 10.17605/OSF.IO/MEUGZ).

2.1. Search Strategy

The electronic databases IEEE Xplore, Scopus, Web of Science, EBSCOHost, and PubMed were searched on 6 May 2025 (between 18:31 PM and 19:56 PM, GMT-5) to identify evidence related to the validity of markerless motion capture systems in human–object interaction tasks, published in scientific journal articles. The search strategy and keywords were defined according to previous systematic reviews evaluating the validity, reliability, and clinical applications of markerless motion capture systems [4,5,6]. The search strategy used in the databases (

Table 1. PECO framework with extracted keywords related to markerless motion capture validation.

P	E	C	O
human	markerless	gold standard	kinematics
able-bodied	marker-less	marker-based	valid *
patient	markerless motion capture	3D marker-based motion analysis	concurrent validity
clinical	markerless motion capture technology	optoelectronic system	accuracy
load carriage	dynamic movement capture	Vicon	error
body-borne load	body scan	OptiTrack	correlation
backpack	pose estimation	Qualisys	concordance
walker	kinect	manual tracking	comparison
crutches	deeplabcut	reference
cane	openpose
assisted-gait	alphapose
handling	Theia3D
	neural network

Abbreviations: PECO = Population, Exposure, Comparison, and Outcome; * = wildcard term.

) was performed by combining keywords and medical subject headings (MeSH) to make the search as exhaustive as possible. Keywords that did not yield any results were discarded after the pilot test. Searches were restricted to the title, abstract, and keyword fields in all the indicated databases.

2.2. Eligibility Criteria

This review only considered original studies that met the following conditions: (i) peer-reviewed scientific articles; (ii) full text available in English language and using a PICOS/PECOS framework [32]; (iii) human participants; (iv) reported at least one kinematic-derived outcome (e.g., gait speed, joint angles); (v) evaluated a markerless motion capture system that was directly compared to a reference system; and (vi) explicitly stated or allowed inference of ethics committee approval [33]. Furthermore, eligible studies had to include participants who passively or actively interacted with a physical object during data collection (e.g., lifting, carrying, manipulating, or moving under external load). No restrictions were imposed on study design or publication date (articles were searched from inception of the database to the end date of the search).

2.3. Study Selection Process

First, all duplicate records were removed using the “Duplicate Items” function associated with the free bibliographic software Zotero (v.7.0.15; Corporation for Digital Scholarship, Vienna, VA USA). The lead author then independently screened the titles, abstracts and keywords of all items retrieved through database searches, applying the pre-specified eligibility criteria in a systematic approach. Exclusions were made when records (i) involved robotic systems; (ii) were conference abstracts, conference proceedings, book chapters or other type of grey literature [16,34,35]; (iii) were identified as retracted publications; (iv) evaluated just the feasibility of a markerless system without validation against a reference method; or (v) described experiments where the authors did not include either evidence of ethical approval or justification for the exemption in the article text [36,37].

2.4. Data Extraction and Evidence Synthesis

For each study extraction, we systematically retained data for the following items: (i) reference information (authors, year of publication, funding source); (ii) sample and setting (country/region, number of participants, gender, age categories (child $\le 12$ years, adolescent 13–17 years, adult 18–59 years and older adult $\ge 60\%$ years)); (iii) task protocol, task location, task instructions, object descriptions, object interactions (e.g., lifting, carrying, reaching, transporting (i.e., locomotion with external load), sporting gestures); (iv) kinematic system, collection method (system/platform, number of cameras/sensors, type of cameras/sensors, and sampling frequency), reference systems used for validation, and any other instrumentation used in conjunction with the kinematic system (e.g., force plates, EMG, inertial measurement units (IMUs)); (v) kinematic outputs, other outcomes and statistical comparisons (e.g., RMSE/RMSD, ICC, correlations, Bland–Altman, TOST, ANOVA) in relation to the main purpose of the present systematic review; and (vi) level/category of evidence given for best-evidence synthesis (see below). When required information was incomplete, or unclear, the corresponding author was contacted once.

The methodological quality of the included studies was assessed by using the SURE critical appraisal checklist for cross-sectional studies (Specialist Unit for Review Evidence, 2018) [38,39,40]. The checklist consists of 12 items that address aspects such as study design, outlines the research question, setting and recruitment, participants, sampling and characteristics, appropriateness of exposure and outcomes, sample size rationale, statistical methods, eligibility reporting, results were described, declaration of conflicts of interest, and acknowledgment of limitations.

Each item was scored based on a 3-point scale (2 = criterion met; 1 = unclear; 0 = not met) to give a total score out of 12, which was then converted to percentage scores. The sum of items ($\mathsf{\Sigma }$) was calculated for each article and transformed into a quality score percentage $QS={\displaystyle \frac{\mathsf{\Sigma }}{24}}\times 100$. According to the adopted threshold, studies with a score < 75% were classified as low quality, while those with a score ≥ 75% were considered high quality [41]. Two authors independently completed the checklist (NU and RC) on all papers included. When a consensus could not be reached, a third (senior) author (LV) was then consulted until a consensus was reached. Best-evidence synthesis was conducted with regard to task type (e.g., lifting, loaded gait, reaching/squats, sport-specific gestures) [42] and reported the strength of evidence based on predefined criteria: (i) the consistency of research results across at least two independent studies examining a similar task (≥75% of studies reporting results showing the same direction), and (ii) the methodology quality rating determined by the SURE checklist. Evidence was deemed strong when the consistent findings were all from high-quality studies, moderate when consistent findings included at least one high-quality study, conflicting when there were inconsistent findings between studies, and limited when there was only one study. The threshold to consider included evidence as statistically significant for the current systematic review was p ≤ 0.05 if not otherwise noted.

3. Results

In total, 3713 studies were initially found in the systematic search across the five databases (Figure 1). After discarding duplicates, 1844 titles were left as unique records. At the initial screening, 339 articles were excluded for low relevance. The titles, abstracts, and keywords were screened for 1505 articles, of which 1443 were excluded for non-compliance with the inclusion criteria previously established in methods.

Then, 62 studies were screened for full-text review for eligibility. Of them, 47 studies were excluded for the following reasons: robotic systems (n = 2), feasibility reports, protocol only (n = 2), no human participants (n = 8), no object interaction (n = 8), no markerless systems (n = 19), no reference system (n = 4), a language other than English (n = 1), no peer reviewed article (n = 1), or absence of reporting kinematic measures (n = 2).

Finally, 15 studies [8,9,10,11,12,13,24,25,26,27,43,44,45,46,47] met all the required criteria and were included in the qualitative synthesis. Reasons for exclusion in the final eligibility stage included, for example, studies that were not in English, did not use a markerless system, or lacked a reference system comparison.

3.1. Characteristics of the Included Studies

The main methodological aspects of the included studies are summarized in

Table 2. Summary of included studies on markerless motion capture validity in human–object interaction tasks.

Reference	Metrics Collected	Markerless System	Reference System	Experimental Protocol	Statistical Treatment
Bonakdar et al., 2025 [12]	Joint angles (back, knee, shoulder, elbow), JRF, REBA	PoseChecker (ML-OMC)	MB-OMC (Vicon), IMUs, force plates	Lift 12.7 kg box floor → pelvis; recordings with markers, IMUs, force plates, iPhone.	Correlations; RMSE, normalized RMSE; REBA vs. expert estimates.
Mehrizi et al., 2018 [24]	3D joint angles (hip, knee, ankle, lumbar, shoulder, elbow), joint positions	Modified Twin Gaussian Process + 2 camcorders	Motion Analysis Corp. (marker-based, 45 markers)	Three symmetric lifts (floor–knuckle, knuckle–shoulder, floor–shoulder) with 10 kg box.	Euclidean distance; joint-angle diff (mean ± SD); paired t-tests.
Ripic et al., 2022 [8]	Spatiotemporal gait (speed, step/stride time, step length, etc.)	KinaTrax (AI-based)	SMART-DX (BTS) + force plates	5 m walkway, self-selected speed, 3 trials; synchronous MB and ML; DLS emphasized.	ICC(2,k), Lin’s CCC; Bland–Altman; paired tests; agreement levels (poor–excellent).
Perrott et al., 2017 [44]	13 joint angles (trunk, pelvis, hip/knee/ankle flexion/rotation)	Organic Motion	Vicon (Helen Hayes model)	(1) Knee Flexion Test with stick; (2) Single-Leg Squat; non-simultaneous sessions.	Paired t/Wilcoxon; Pearson/Spearman; start → peak change and peak-angle comparisons.
Wren et al., 2023 [9]	Lower-limb kinematics; RMSD, RMSDoffset, mean diff over gait cycle	Theia3D	Vicon Nexus (Plug-in-Gait)	Pediatric gait; concurrent capture; subgroups: orthoses, deformities, devices.	RMSD, RMSDoffset, mean diff; subgroup analyses; visual waveforms.
Torvinen et al., 2024 [10]	Joint-center 3D distance, joint-vector angles (elbow, shoulder, hip, knee, ankle); RMSE, ICC, r	DeepLabCut (custom)	Vicon Nexus 2.8.1 (8 cams)	Experienced skiers; G1/G3 skating on treadmill with poles; ML model trained on separate cohort.	Bland–Altman; RMSE; Pearson r; ICC; mean ± SD; LoA; $p\le 0.05$.
Wagh et al., 2024 [47]	2D index-finger trajectories; RMSE vs. touchscreen; frame- rate effect	MediaPipe Hands	Touchscreen (Surface Pro 7)	Trace animated shapes on touchscreen; camera capture (GoPro); resampling, normalization, Procrustes.	RMSE; TOST for equivalence; t-tests across 30/60/120 FPS; $\alpha <0.05$.
Gupta et al., 2016 [45]	Stride length/width, height of earlobe; with/without backpack	Kinect V2	Clinical manual method	7 m walkway; pre/post ergonomic repacking (AOTA); with vs. without backpack.	Paired t-tests; Bland–Altman; Pearson r; $R^2$; % error; linear/cubic regressions.
Steinebach et al., 2020 [25]	Shoulder (abd, flex) and elbow (flex) angles; static and dynamic; with/without box	Kinect V2 × 2 + Captiv IMUs	Goniometer (static), angle scale (dynamic)	(1) Static postures; (2) movements w/o objects; (3) with box (occlusion).	MAE, RMSE; Pearson/Spearman; Wilcoxon; Bland–Altman; $\alpha =0.05$.
Scano et al., 2020 [27]	3D joint angles (shoulder, elbow, wrist, trunk); 11 DoF; angular distance; test–retest	Kinect V2	Vicon Vero (10 IR cams)	Seated right-arm pointing and workspace exploration across 3 sectors.	RMSE; angular distance; ICC (retest); two-way ANOVA (DoF × sector), Tukey HSD; Bland–Altman; $\alpha =0.05$.
Mehrizi et al., 2019 [46]	3D body pose; L5/S1 force and moment (kinetics); peaks & time series	DNN + 2 RGB cams	Motion Analysis Corp. (marker-based)	Sym/asym lifting of 10 kg crate across 9 conditions (3 heights × 3 end angles); synchronized capture.	RMSE; R coefficient; ICC (peaks); Bland–Altman; ICC $>0.99$ for peak moments/forces.
Coll et al., 2025 [11]	Joint angles (hip, knee, ankle); joint reaction forces; EMG timing	Theia3D (7 RGB cams)	Vicon (11 IR cams, 120 Hz) + force plates + EMG	Walk/run under 4 military loads (5–41 kg) over 6 m; real rifle; OpenSim modeling.	RMSE; Pearson r; Bland–Altman LOA; RM-ANOVA (joint × load), Tukey; forces to body-weight (×BW).
Remedios & Fischer, 2021 [26]	Peak joint angles (knee, trunk, shoulder); shoulder abd; FSL; COG-to-load; hip–knee MARP	Wrnch AI (2D, sagittal)	Vicon 3D (60 Hz)	Box lifts floor → waist under 3 loads; simultaneous 2D vs. 3D posture/balance/ coordination.	Bland–Altman; LOA; Shapiro–Wilk; outlier checks; descriptives.
Kwolek et al., 2019 [43]	3D joint trajectories; Euclidean joint-position error (ankle, knee, hip, etc.)	Not reported (8 RGB cams)	Vicon MX-T40 (10 cams)	166 sequences; walking under 4 conditions (normal, clothes change, backpack, varied gait); RGB 25 Hz, Vicon 100 Hz.	Mean joint error (cm) per joint; no inferential stats (dataset benchmarking).
Bae et al., 2024 [13]	3D joint angles (shoulder, hip, knee); ROM; time-series and peaks; test–retest	Ergo system (18 RGB + DL)	Qualisys (8 cams); 36 markers (Helen Hayes)	3 trials of overhead squat with 120 cm dowel; concurrent capture.	ICC(2,1), ICC(3,1); RMSE; Bland–Altman; CV; linear regression

Abbreviations: RMSE = root mean square error; RMSD = root mean square deviation; REBA = Rapid Entire Body Assessment; r = Pearson correlation coefficient; R² = coefficient of determination; Bland–Altman = Bland–Altman plot/limits of agreement; TOST = two one-sided tests; ANOVA = analysis of variance; ICC = intraclass correlation coefficient.

. The total number of participants across all studies was 316 (median: 21 participants per study). Regarding sex, three studies included exclusively males (20%), while twelve studies (80%) included both men and women. As for age groups, most studies (12, equivalent to 80%) were conducted with adult samples. Only one study included children (≤12 years), one included adolescents, and one included older adults. In one study, information about the age group was not clearly reported.

The experimental protocols included (i) symmetrical or asymmetrical lifting of boxes (from 10 to 12.7 kg) (five studies; [12,24,25,26,46]), (ii) gait and locomotion tests under different conditions, including manipulation of external loads such as backpacks or military equipment (five studies; [8,9,11,43,45]), (iii) squat and knee flexion tests (two studies; [13,44]), (iv) reaching and manipulation tasks (two studies; [27,47]), and (v) technical gestures such as treadmill skating [10]. In most studies, the main task was clearly reported, although only in 38% ([11,13,25,26,46]) was the duration or exact number of repetitions of the tests specified. Figure 2 shows the distribution of included studies published over the years. The older study on the subject found in the current review was published in 2016 [45]. Between 2016 and 2018 and 2021 and 2023, one study was published per year; increases were observed between 2019 and 2020 (two studies published per year) and more recently in 2024 (three studies published; [10,13,47]).

The metrics collected predominantly included three-dimensional joint angles of the upper and lower limbs (shoulder, elbow, wrist, hip, knee, ankle) (present in at least 9 studies: [9,10,11,12,13,24,27,44,46]). Spatiotemporal gait parameters (speed, stride/step time and length) were also analyzed (three papers; ([8,43,45])), as well as joint forces and moments (including L5/S1 forces) ([11,12,46]), muscle activation via EMG [11], and functional or ergonomic metrics such as REBA scores [12]. All of the included studies reported at least one kinematic metric, while 54% included additional kinetic or functional variables. Regarding markerless motion capture systems, the following solutions were used across the included studies: artificial intelligence and deep learning algorithms (PoseChecker, DeepLabCut, Theia3D, MediaPipe Hands), commercial systems such as Microsoft Kinect V2 (used in at least 4 studies), Organic Motion, Wrnch AI, and the Ergo system, as well as custom platforms in some cases.

Deep learning-based platforms, such as the Ergo system [13], demonstrated excellent concurrent validity and high test–retest reliability in functional movements like the overhead squat, with RMSE values ranging from 2 to 6° and ICCs up to 1.0. The reference systems were predominantly optical marker-based (Vicon, Yarnton, UK, Qualisys Gothenburg, Sweden, Motion Analysis Corp., Rohnert Park, CA), employed in 100% of the studies for direct comparison. In more comprehensive protocols, these systems were often accompanied by IMUs, force plates, or EMG. Kinect was the most frequently used markerless system, appearing in 3 of the 15 included studies [25,27,45]. These studies assessed gait, overhead lifting, and upper-limb reaching tasks. Although methodological differences exist, all three reported acceptable to high validity results when compared to reference systems, including RMSE values below 10° or 10 mm, and ICCs above 0.85, supporting Kinect’s suitability for basic clinical and ergonomic assessments.

The number and arrangement of sensors/cameras varied considerably: from simple setups with a single camera (MediaPipe Hands in [47]) to multi-camera systems of 7, 8, 10, 11, and up to 18 RGB or infrared cameras [10,11,13]. The anatomical placement of markers was described in detail in 92% of the studies, covering the pelvis, hip, thigh, leg, ankle, foot, trunk, shoulder, elbow, wrist, hand, and head, depending on the experimental objective.

As for acquisition frequencies, this information was explicitly reported in only some studies, with typical values ranging from 25 Hz [43], 60 Hz, 100 Hz, 120 Hz (Vicon in [11]), and up to 200 Hz (Qualisys in [13]). In approximately 31% of the studies, the acquisition frequency was not specified for all systems used or was only indicated for the marker-based system.

The statistical strategies employed almost always included root mean square error (RMSE), Pearson or Spearman correlation coefficients, intraclass correlation coefficients (ICC), Bland–Altman analysis, paired t-tests, and in several cases, repeated-measures ANOVA and equivalence tests. All of the studies applied at least one quantitative comparison between systems, strengthening the validity of their results. Three studies used 2D (20%; [25,26,47]) while twelve used 3D analysis (80%; [8,9,10,11,12,13,24,27,43,44,45,46]) regarding the tracking method.

3.2. Measurement Properties—Overview

The main results related to the measurement properties of markerless motion capture systems for human–object interaction tasks are summarized in

Table 3. Summary of validity results for included markerless motion capture studies addressing human–object interaction tasks.

Reference	Sample (N, Sex, Age)	Markerless System	Object	Validity Results	Summary
Bonakdar et al., 2025 [12]	$N=8$; 4M/4F; $25\pm 3$ y	PoseChecker (ML-OMC)	Box (11.3 kg); ergonomic load; repetitive lifts	RMSE $=6.5$–$12.9°$; $r=0.81$–$0.95$	Authors report good validity; posture scoring aligns with experts.
Mehrizi et al., 2018 [24]	$N=12$; all male; $47.5\pm 11.3$ y	2 RGB cams + ML model	Box (10 kg); lifts with postural constraints	RMSE $=5$–$12°$	Valid for estimating spinal joint loads during lifting.
Ripic et al., 2022 [8]	$N=22$; 10M/12F; $22.72\pm 3.49$ y	KinaTrax (AI skeletal)	Force plate (5 m walkway); stepped during walking	ICC ≈ 0.87–0.93	Valid for spatiotemporal gait analysis.
Perrott et al., 2017 [44]	$N=20$; 10M/10F; median 28.1 (22–40)	Organic Motion	PVC stick + squat box; used during squats	RMSD $=5$–$7°$; very weak-to-strong correlations	Acceptable validity for squats and reach movements.
Wren et al., 2023 [9]	$N=36$; 20M/16F; 2–25 y (mixed)	Theia3D	Pediatric orthoses, walkers; worn during gait	RMSD and offset $\approx 5°$	Valid for pediatric gait with assistive devices.
Torvinen et al., 2024 [10]	$N=10$; 5M/5F; F: $21.0\pm 3.1$ y; M: $24.4\pm 10.7$ y	DeepLabCut (4 RGB)	Roller skis + poles; treadmill skating	RMSE $=2.9$–$4.6°$; ICC $>0.90$; r = 0.82–0.97	Provided valid kinematics in treadmill skiing.
Wagh et al., 2024 [47]	$N=10$; 9F/1M; $19.5\pm 1.3$ y	MediaPipe Hands	Touchscreen panel; fingertip tracing of shapes	RMSE $=4.1$–$6.2$ mm	Valid for capturing 2D fingertip trajectories.
Gupta et al., 2016 [45]	$N=60$; 60M; $11.77\pm 1.52$ y	Kinect V2	Backpack (∼10% BW); worn during 7 m walkway gait	strong correlation (r1 and 2 > 0.90)	Limited but acceptable validity in children’s gait.
Steinebach et al., 2020 [25]	$N=12$; 7M/5F; $23.8\pm 2.6$ y	Kinect V2 + Captiv IMUs	Box; overhead reach/ lift task	RMSE $=6$–$8°$; ICC $>0.85$; r $>0.93$	Valid for upper-limb angles during lifting.
Scano et al., 2020 [27]	$N=15$; 8F/7M; $24\pm 3$ y	Kinect V2	Tabletop targets; seated pointing	RMSE $=9.6°$; ICC $>0.90$	Accurate and repeatable for seated pointing tasks.
Mehrizi et al., 2019 [46]	$N=12$; 12M; $47.5\pm 11.3$ y	2 RGB cams + DNN	Box (10 kg); simulated lift postures	RMSE $=6.7$–$11.3°$; ICC $>0.89$; R $>0.80$	Valid for estimating lumbar loads during lifting.
Coll et al., 2025 [11]	$N=16$; 8F/8M; $30\pm 12$ y	Theia3D	Military gear (helmet, vest, rifle); worn during tasks	ICC $>0.85$; $r>0.80$; significant ANOVA effects	Robust validity in high-load military simulations.
Remedios & Fischer, 2021 [26]	$N=20$; 11M/9F; M: $21.6\pm 3.1$ y; F: $21\pm 1.2$ y	Wrnch AI (2D)	Box; lifts to 3 levels for posture scoring	Mean differences $1.9$–$22.1°$ between 2D MMC and 3D reference; Bland–Altman LoA $\approx \pm 20°$	Valid for ergonomic posture classification.
Kwolek et al., 2019 [43]	$N=32$; 10F/22M; age NR	8 RGB cams + ML triangulation	Backpack (7 trials); worn during walking	Rank-1 accuracy $=96\%$	High gait-recognition accuracy; not a biomechanical validity study.
Bae et al., 2024 [13]	$N=31$; 22M/9F; $32.2\pm 5.9$ y	Ergo system (DL, multicam)	Overhead squat; full- body kinematics	RMSE $=2.3$–$6.3°$; ICC $=0.75$–$1.0$; $R^2$ = 0.88–0.99	Excellent concurrent validity and high test–retest reliability.

Abbreviations: RMSE = root mean square error; RMSD = root mean square deviation; r = Pearson correlation coefficient; R² = coefficient of determination; Bland–Altman = Bland–Altman plot/limits of agreement; ANOVA = analysis of variance; ICC = intraclass correlation coefficient. Object = Characteristic/Type of interaction.

. Overall, the authors claimed validity of markerless systems in 13 of the 15 included studies (87%). For example, high correlations (corr = 0.81–0.95) and acceptable root mean square errors (RMSE = 5–23°) were observed for estimating joint angles and ergonomic loads when comparing PoseChecker with reference systems [12]. Similar results were reported with RGB camera-based models and machine learning [25,46], with RMSE between 5 and 12° and high correlations for lumbar spine loads. Other studies using commercial and artificial intelligence systems, such as KinaTrax, Theia3D, DeepLabCut, or MediaPipe Hands, or the Ergo platform [13] also showed good agreement with traditional marker-based systems for gait analysis, reaching and tracing tasks, and ergonomic assessments [8,9,10,47]. On the contrary, 2 studies did not support the validity of markerless systems. In these, gait tasks were studied either with limited validation [45] or distinguishing biometric classification rather than biomechanical [43].

In reference to the manipulated objects, most studies analyzed tasks involving only rigid tools/loads (11/15 studies, 73%; [8,9,10,12,13,24,25,26,27,46,47]), whereas only four studies (27%) included non-rigid or mixed rigid/non-rigid objects such as hurdles, backpacks, or tactical vests [11,43,44,45]. Among the studies involving only rigid tools/loads, validity was supported in 8/11 studies (73%) and not confirmed in 3/11 (27%). In contrast, for tasks involving non-rigid or mixed rigid/non-rigid objects (e.g., hurdles, backpacks, tactical vests), only 2/4 studies (50%) supported the validity of MMC, while the remaining 2/4 (50%) reported insufficient agreement.

Regarding reliability and accuracy, the results were also generally favorable, although fewer studies explicitly evaluated these properties. For example, ICCs > 0.85 were reported for dynamic tasks (box lifting, gait, squats), both in within-session protocols and under controlled conditions [10,25]. However, between-session reliability and long-term reproducibility have been poorly explored, and in some cases only within-session comparisons were documented.

In short, in the included studies using correlation analysis, six (86%) demonstrated high associations [10,11,12,13,25,45] while only one reported mixed results (i.e., weak-to-strong correlations; [44]). ICCs reported were also generally high, ranging from 0.75 to 1 across all studies which calculated such metric [8,10,11,13,25,27,46]. On the other hand, RMSD/RMSE values notably varied across studies from approximately 2° [13] to 13° [12] (see Table 3).

3.3. Methodological Quality and Risk of Bias in the Included Studies

The overall methodological quality outcomes for the included studies, as well as scores for each of the 12-item scale, are summarized in

Table 4. Quality assessment of included studies (Q1–Q12, sum score $\mathsf{\Sigma }$, and quality score percentage QS).

Reference	Q1	Q2	Q3	Q4	Q5	Q6	Q7	Q8	Q9	Q10	Q11	Q12	$\mathbf{\Sigma }$	QS
Bae et al., 2024 [13]	2	2	1	2	2	2	2	2	2	2	2	2	23	96
Bonakdar et al., 2025 [12]	2	2	1	1	2	2	0	2	2	2	2	2	20	83
Coll et al., 2025 [11]	2	2	1	1	2	2	0	2	2	2	2	2	20	83
Gupta et al., 2016 [45]	2	2	1	2	0	1	0	1	0	1	2	0	12	50
Kwolek et al., 2019 [43]	2	2	1	1	0	2	0	2	2	2	2	0	16	67
Mehrizi et al., 2018 [24]	2	0	1	1	2	2	0	2	0	2	2	2	16	67
Mehrizi et al., 2019 [46]	2	2	1	1	2	2	0	2	1	2	2	2	19	79
Perrott et al., 2017 [44]	2	2	1	1	0	1	0	2	0	2	2	2	15	63
Remedios & Fischer, 2021 [26]	2	2	1	2	2	2	2	2	2	2	2	2	23	96
Ripic et al., 2022 [8]	2	2	1	0	2	2	0	2	1	2	2	2	18	75
Scano et al., 2020 [27]	2	0	1	1	2	2	0	2	1	2	2	2	17	71
Steinebach et al., 2020 [25]	2	2	1	1	2	2	0	2	1	2	2	2	19	79
Torvinen et al., 2024 [10]	2	2	1	1	2	2	0	2	2	2	2	2	20	83
Wagh et al., 2024 [47]	2	2	1	1	0	2	0	2	2	2	2	2	18	75
Wren et al., 2023 [9]	2	2	1	2	2	2	0	2	2	2	2	2	21	88
Mean	2.00	1.73	1.00	1.20	1.47	1.87	0.27	1.93	1.33	1.93	2.00	1.73	18.47	76.94
SD	0.00	0.70	0.00	0.56	0.92	0.35	0.70	0.26	0.82	0.26	0.00	0.70	2.97	12.39

. Sample size was the most frequently low-scoring aspect, with 100% of the studies scoring suboptimal on this criterion (Q3). Furthermore, other relevant methodological elements, such as the absence of intraclass reliability (ICC) analysis and the lack of correlation or area under the curve (AUC) analysis, were also identified as recurring missing items in part of the included studies, i.e., respectively, 46% and 38% of the studies lacked or were unclear in reporting these analyses. Overall, most studies achieved moderate to high overall methodological quality, with an average score of 77% (ranged between 50% and 96%). The highest-quality studies reached a 96% score [13,26], while the lowest score corresponded to 50% [45].

3.4. Evidence Synthesis

The best-evidence synthesis tool was adopted in an attempt to depict the results according to the group of tasks analyzed in included studies (see

Table 5. Evidence synthesis regarding the concurrent validity of markerless tracking to evaluate kinematics in human–object interaction tasks.

Group of Tasks	Methodological Quality	Validity Confirmed	Validity not Confirmed	Judgement Evidence
Lifting	High-quality	Bonakdar et al., 2025 [12]; Mehrizi et al., 2019 [46]; Steinebach et al., 2020 [25]; Remedios & Fischer, 2021 [26]	—	Strong evidence
	Low-quality	Mehrizi et al., 2018 [24]	—
Locomotion	High-quality	Coll et al., 2025 [11]	Wren et al., 2023 [9]	Conflicting evidence
	Low-quality	—	Gupta et al., 2016 [45]; Ripic et al., 2022 [8]; Kwolek et al., 2019 [43]
Squat/knee flexion	High-quality	Bae et al., 2024 [13]	—	Moderate evidence
	Low-quality	Perrott et al., 2017 [44]	—
Reaching/ manipulation	High-quality	Torvinen et al., 2024 [10]	—	Conflicting evidence
	Low-quality	Wagh et al., 2024 [47]	Scano et al., 2020 [27]

). After applying the criteria to classify the level of evidence, the following results were obtained: (a) lifting tasks had strong evidence for the use of markerless systems, as we found that the results were consistent across five studies (four high quality and one low quality) [12,24,25,26,46]. These studies reported strong correlation coefficients and acceptable error margins (RMSE = 5°–23°); (b) for gait and locomotor tasks, especially load manipulation tasks, we found conflicting evidence across 5 studies [8,9,11,43,45]; (c) squat and total knee flexion tests had moderate evidence, as the results were consistent in two studies [13,44] of different quality; and (d) reaching/manipulation tasks presented conflicting evidence across studies [10,27,47]. Overall, the most robust (consensus) evidence from the content provided was related to lifting tasks and less consistent among gait, reaching, and sport-specific complex movements.

4. Discussion

The main objective of the current review was to systematically evaluate the evidence on the measurement error assessments of markerless motion capture systems, specifically for movement analysis involving human–object interaction. A total of 15 articles met all eligibility criteria and therefore were reviewed and their evidence collated. The key findings are as follows: (1) there is strong evidence supporting the validity of markerless systems during lifting tasks; (2) only moderate or conflicting evidence was found for gait, squats, reaching, and complex sports-related gestural movement tasks; (3) the most frequently used markerless system was the Kinect V2, used in four studies which all reported acceptable to high validity for basic movements despite methodological variability; and (4) overall, most studies achieved moderate to high methodological quality (average score 77%), though recurring weaknesses were identified, particularly in sample size justification, reporting of reliability analyses, and the lack of use of some statistical methods (e.g., between-session ICCs). The following paragraphs present the main findings of this review, contrast them with those of previous systematic and narrative reviews on markerless motion capture systems, and offer interpretative insights. Based on these comparisons, recommendations are proposed to guide future research and clinical application.

Our review found strong evidence supporting the validity of markerless motion capture systems for symmetrical/asymmetrical lifting tasks. In at least five studies, including four high-quality studies, markerless systems showed high correlation coefficients and acceptable errors in estimating joint angles and ergonomic loads during box lifting. This is consistent with the previous literature [48,49], which suggests that repetitive, controlled, and relatively simple tasks, like weightlifting, are more suitable for markerless tracking, as they likely involve lower complexity and reduced movement occlusion. Most studies used several AI-based systems or multi-camera configurations to improve measurement accuracy and robustness, for example, PoseChecker [12], Theia3D [9,11], DeepLabCut with 4 RGB cameras [10], and an 8-camera setup with triangulation [43]. These configurations may have played a role to improve joint reconstruction and measurement fidelity. However, due to the fact that some types of multiple-camera systems, even if they are of the same model, may present phase differences between them [50], it cannot be discarded that this may have an impact on the levels of error reported for the video-based MMC tracking systems.

Overall, we did not observe that errors due to differences in joint center estimation or brief occlusion affected the construct validity results, except for two reports that presented methodological limitations [43,45]. Taken together, these findings suggest that markerless systems could be adopted as an alternative monitoring tool in ergonomic and occupational health assessments related to lifting objects while encouraging future research into more complex or real-world lifting tasks. Regarding the evidence for gait/locomotion tasks including the manipulation of external loads, the evidence supporting markerless tracking application was more variable and, in some cases, even contradictory. Across all studies on gait and locomotion tasks, varying levels of agreement with marker-based systems were observed, especially with healthy adults performing in controlled settings or stable environments like treadmills, walking or laboratory settings. Some studies reported good agreement with marker-based systems [8,9,45] while others, particularly studies in children or during tasks designed to measure load adaptation, such as backpack or military load, demonstrated limited or partial validity [11,43,45]. Furthermore, for squats, reaches, and treadmill skating, only a few studies provided moderate evidence for the use of markerless systems in those settings, only some of which had limited methodological quality [10,44,47]. These trends support previous findings [30,48] suggesting that markerless systems still struggle to achieve the unrestricted, multi-joint, or rapid movements intended to be quantified. These inconsistencies may arise not only from differences in execution or participant characteristics but also from intrinsic limitations in the way that current markerless systems characterize complex movements. While previously published reviews (e.g., [4,30,48]) have shown strong validity, they have focused on tasks associated with unconstrained, unloaded movement. In contrast, we have focused on the task of human–object interaction and therefore consider behaviors such as load carriage or manipulation of external objects, which might bring with it additional aspects of biomechanical complexity or marker occlusions that would limit the algorithm’s ability to be able to track joint trajectories. This may explain the lower or inconsistent validity outcomes we observed in more “real-world” tasks, for example, gait under load or military marching. Indeed, one can argue that various algorithms used for object recognition in markerless tracking systems may not have been originally designed for the specific purpose of analyzing human movement from a mechanical point of view at segments level, and this could contribute to the (still) lack of acceptable validity in some cases [1]. Therefore, it is recommended that markerless systems be used with caution for these tasks, and additional high-quality studies are needed to confirm their validity.

Furthermore, several studies examined specific or challenging conditions that may influence system performance. These included external loads such as school backpacks [45], military packs [11], lifting technique variations [25], and orthoses, deformities, or assistive devices in children [9]. Some studies additionally investigated intra- and inter-subject repeatability and gesture execution in varied contexts. While these contexts augment the ecological validity of markerless assessments, they also added additional variability. As, for example, the results of a lower percentage of studies supporting the validity of MMC in conditions of non-rigid or rigid/non-rigid objects suggests that MMC systems tend to perform more consistently in tasks with rigid objects, whereas non-rigid or deformable loads still pose important challenges for accurate kinematic estimation.

The Kinect V2 was the most often used commercial markerless system used by the studies included in this review. However, validity was not consistent across studies. While there were studies assessing simple gait or upper-limb motion, with several studies reporting acceptable agreement with marker-based systems [45], there were studies where more complicated motions like lifting, overhead contexts, or gait under load were only partially or limited valid [25,43]. Trainers and companies using Kinect V2 will be drawn to its lower cost, ease of access, and ease of set-up, but this does not reflect player performance when the demand of the task is greater. While markerless systems have less value for biomechanical research, studies using enhanced markerless systems or multi-camera arrays such as Theia3D, DeepLabCut, or the Ergo platform [13] report better validity in complex or dynamic tasks highlighting not only the selection but also experimental design aspects which impact the validity.

For research and practice applications requiring high-quality measurement and/or complex movement, advanced markerless systems should be considered, and researchers/ practitioners should be transparent about task constraints and system limitations. Consistent with the results described, most authors concluded that the markerless systems they evaluated were valid and accurate (see Table 3) and, in some cases, reliable for capturing kinematic and functional parameters. Authors particularly emphasized portability, reduced cost, and avoidance of physical markers in tasks where there are practical reasons for markerless systems to be favored over competitive systems [9,25,26,27,45]. Specifically, 87% (13 studies) of the studies we reviewed concluded markerless motion capture systems were valid for at least one of the evaluated tasks. In contrast, 13% (2 studies) reported partial or insufficient validity, typically for complex, uncontrolled movements. Findings suggest an emerging consensus on the value of markerless systems, albeit they still need to be interpreted concerning the task and methodology.

In general, the overall methodological quality of the included studies was moderate to high (mean quality score 77%, range: 50–96%), indicating progress towards a rigorous approach to research on the topic. However, many studies continued to exhibit recurring limitations, particularly regarding sample size justification, reliability analysis (ICC, AUC), and standardizing measurement protocols. This is consistent with previous reviews describing significant and ongoing deficiencies in statistical and methodological power, mainly due to deficiencies in sample size and reporting of advanced statistical analyses [30,48,51]. Additionally, in approximately 31% of the studies, the acquisition frequency was not provided for one or more of the systems, or the acquisition frequency was provided for the marker-based system but not the markerless system. This is a setback, as it did not allow for the assurance of fair comparisons. Although the markerless system probably used the same acquisition frequency in most cases where the information was unclear or omitted, a complete description relating to that methodological feature is necessary. Future studies should document the acquisition frequencies of both systems and, ideally, adjust them to the same sampling frequency before beginning the data collection to allow direct comparisons of motion data. Critical developments to ameliorate these issues are essential to strengthening the comparability and confidence in future research initiatives.

The reviewed studies reveal a lot of variance in their methods; however, they approached their studies for the validation of markerless systems using functional, ergonomic, or sporting tasks. The existing markerless literature regarding human–object interactions has several significant limitations. The majority of the studies employed small or convenience samples, which restrict statistical power and generalizability of findings. There are substantial differences in the protocols, tasks, and outcome measures that make it difficult to compare the studies, and as a result, meta-analyses are impossible. There were very few clinical populations or in situ contexts reported in the studies, and when there were, most studies were confined to a laboratory. This effectively restricted the applicability of the results to real-world contexts. There were also multiple studies that underreported the reliability statistics and failed to provide between-session consistency, and there were also studies that did not report any follow-up assessments, which limits well-evaluated performance. Overall, it is evident that future studies need to enhance methodological transparency, provide adequate statistical methods, justify their sample size, and offer more representative samples of participants. Future development regarding markerless tracking should consider providing user-friendly executable software applications for the provision of real-time data. Also, Neural Radiance Fields (NeRFs) and 3D Gaussian Splatting (3DGS) techniques are tools that may improve occlusion robustness and thus potentially help next-generation markerless systems [52]. Implementation of testing protocols in the real world are warranted as well.

Finally, an important finding of the present systematic review was that across the included studies, correlations between MMC and reference systems were generally high, indicating good agreement as regarding signal waveforms. However, joint-angle errors (RMSE) were more unstable, showing values that would often not be considered negligible in clinical or ergonomic contexts, particularly near end-range of motion or when small angular changes affect risk scores. These findings suggest that current MMC systems can be able to capture the overall waveform of joint kinematics, whereas their precision relative to gold-standard reference systems still requires improvement. It is important to note that some of these differences may be due to the generally distinct definitions of local frame orientations between MMC and marker-based systems [53].

The current systematic review also has limitations, which the authors have acknowledged. Firstly, although a full search strategy was performed, only 15 studies met our strict inclusion criteria, and the results may be limited to the specific studies summarized here. The relatively small number of studies included may be due to the fact that the topic of pose estimation has been researched more intensively since approximately the beginning of the 2020s (see, for example, [54]). Also, based on some previous reports using PRISMA guidelines (e.g., [16,34,35]), grey literature including studies published in conferences were not considered. On the other hand, significant efforts on MMC research and their applications can be found in a range of conferences (e.g., [55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79]), which may have resulted in a possible reduction of strength in the evidence synthesis described in the current review study. Studies on robots were also not considered here (i.e., one of the inclusion criteria was that experiments should have included human participants). However, as there is increasing research involving robots attempting to simulate/recreate human movements (e.g., [80,81,82]), removing studies of such type might lead to a loss of potentially useful information.

In addition, exclusion of studies not published in English, and studies without direct comparisons to marker-based systems, may have resulted in relevant data being omitted from the overall synthesis. Moreover, this synthesis was reliant on the original reports being complete and valid, and any report with writing deficiencies in primary studies may have affected the conclusions drawn here. Finally, our best-evidence synthesis per task category was constrained by the literature; that is, some types of movements remain limited in terms of providing the related literature.

5. Conclusions

The main objective of this review was to evaluate the measurement error of markerless motion capture systems for movement analysis in task assessment involving human–object interaction, using the best evidence synthesis method to collate evidence. Strong evidence existed supporting the validity of markerless tracking application for lifting tasks, especially in a controlled environment. There was moderate evidence for tasks such as squats and knee flexion, but there was some variation across systems and protocols. Conflicted or inconclusive evidence was apparent for reaching and manipulation tasks as well as more complex, sport-specific gestures; likely influenced by variation in protocol design, movement dynamics, and evaluation methods used in each of the studies. There was limited evidence for gait. Finally, it is important to note that MMC generally seems well suited to capturing the form of movement, but its level of precision still requires improvement regarding tasks involving human–object interactions.