Abstract
In keeping with the topic of this special issue, “Breakthroughs in Stroke Rehabilitation: Bridging Engineering, Neuroscience, and Motor Control,” this commentary addresses the recent emergence of computer vision motion capture (CVMC, aka markerless motion capture) with a realistic check-in on its current measurement performance and future utility as a clinical assessment tool. These are heady times. It would be understandable to misinterpret early demonstrations of CVMC in clinical research as a suggestion that this innovative tool is ready for clinical deployment. In reality, benchmarks for CVMC measurements are still being established, and cultivation of targeted clinical practice approaches informed by motion analysis remains aspirational. In this commentary, we reframe the CVMC conversation by first acknowledging the current state of CVMC as a technology still in development. We then consider the path to a long-term goal: targeted stroke rehabilitation in clinical practice informed by the quantification of movement function. In lighthearted spirit, we channel Douglas Adams’ book series, “The Hitchhikers Guide to the Galaxy” (HG2G; Adams, 1979–1992), providing clinician hitchhikers some recommendations to deftly navigate the CVMC landscape on their intergalactic motion capture travels between lab and clinic. And so, just as Ford Prefect, the friendly alien researcher in the HG2G book series would say to reassure new hitchhikers: “You just come along with me and have a good time. The Motion Capture Galaxy's a fun place. You’ll need to have a fish in your ear.” - Ford Prefect (adapted from Adams; Book 1, Chapter 5).
Development of CVMC to identify and track human movement is not new. Progressive work of “hyperintelligent, pan-dimensional beings” (Adams, 1979–1992) over the last four decades (Corazza et al., 2006; Hogg, 1983; Toshev & Szegedy, 2014) has led to the current state of this technology. Yet, it was the recent culmination of large-scale object detection datasets (Lin et al., 2014), open source trained algorithms (Cao et al., 2019), and powerful graphics cards for personal computers that placed these novel tools in the hands of tech-savvy biomechanists and clinical researchers, enabling us to join the CVMC conversation. Today, CVMC development continues in dynamic transition as different hardware solutions are tested, algorithms and training data sets are updated, and measurement performance is interrogated by movement task and study population. As exciting as the potential applications of CVMC may be for stroke rehabilitation, it is prudent we maintain realistic expectations by monitoring the current state of this rapidly evolving technology. We encourage curious, early adopters to explore CVMC with cautious optimism and consider how quantifying functional movements could enable targeted stroke rehabilitation. In homage to Douglas Adams’ last radio broadcast series which considered the integration of technology in both our present and future society, (The Hitchhiker's Guide to the Future, 2001), this commentary is divided into two parts: The Clinician's Guide to the Present contextualizes the current state of CVMC as the next technological development on the motion capture timeline, and The Clinician's Guide to the Future considers a potential paradigm shift in stroke rehabilitation enabled by access to motion analysis in clinical practice.
The Clinician's Guide to the Present
“The story so far:
In the beginning the Motion Capture Lab was created.
This had made a lot of people very happy and been widely regarded as a good move.”
- HG2G Narration (adapted from Adams, 1979–1992; Book 2, Chapter 1).
While biplanar video fluoroscopy is considered the current gold standard in motion capture measurement, due to its limitations and practical constraints, the de facto technique to quantify 3D human movement is marker-based motion capture (MBMC) (Winter, 2009). The availability of professional-grade MBMC systems over the last 30+ years helped establish an infrastructure of laboratories operated by experts across disciplines from engineering to movement science to rehabilitation science. Their work built the knowledge base and methodologies which established MBMC as a trusted measurement tool with standards of precision and accuracy. Undoubtedly, this renaissance period with MBMC has expanded our knowledge of human motion and advanced our understanding of motor dysfunction. MBMC measures are used to determine pathomechanics, track disease progression, document efficacy of a rehabilitation treatment and capture complex behavior patterns difficult to perceive by a well-trained human eye (Marks et al., 2018).
As remarkable as these advances have been, MBMC has its flaws. Motion capture systems remain expensive to acquire and maintain. To be used effectively, they require a controlled laboratory space and dedicated professionals trained in equipment operation, biomechanical models, anatomy, and good old engineering problem solving. Even in places where a motion capture lab is accessible, measurement precision and accuracy remain dependent on correct marker placement. “The major problem – one of the major problems, for there are several – one of the many major problems with MBMC is that affixing reflective markers to anatomical landmarks introduces measurement error due to incorrect placement, skin movement artifact, and underlying adipose tissue which varies across study populations. To summarize: it is a well-known fact that measurements reported using MBMC are, ipso facto, a best estimation of the desired position coordinates. To summarize the summary: anyone who is capable of placing stickers on skin can report inaccurate measures of 3D human movement. To summarize the summary of the summary: people can be a problem.” - HG2G Narration, a playful commentary on the leadership paradox (adapted from Adams, 1979–1992; Book 2, Chapter 28).
Acknowledging this information allows us to appreciate one of the advantages of CVMC: measurement in the absence of reflective markers.
Key to our current interest in the clinical translation of motion capture tools, traditional MBMC also places a burden on the studied individual in terms of travel distance to the lab, time commitment, willingness to be touched and marked, and to say nothing of “acting naturally” while scantily clad in an unfamiliar environment under scrutiny of strangers. Reducing this burden on the patient/study participant is an appealing feature of CVMC applications, especially for use in clinical research and practice.
“ “Many many millions of years ago … hyper-intelligent pan-dimensional beings … got so fed up with the constant bickering about the meaning of life … (Why are people born? Why do they die? Why do they want to spend so much of the intervening time wearing digital watches?) … that they decided to sit down and solve their problems once and for all.” - HG2G Narration (Adams, 1979–1992; Book 1, Chapter 25).
So, they built a supercomputer named Deep Thought which, after running a 7.5-million-year research program, completed its calculations and reported the “Answer to Life's Ultimate Question” to be: “’Forty-two’, said Deep Thought, with infinite majesty and calm.” - Deep Thought (Adams, 1979–1992; Book 1, Chapter 27).
Unlike the anticlimactic and meaningless “Answer” (42!) put forward in HG2G, there is currently no “Answer to the Ultimate Question of Everything” markerless (Adams, 1979–1992; Book 1, Chapter 25). Multiple promising CVMC systems are actively being tested and benchmarked for implementation in practical applications and clinical research. Yet no single CVMC system ticks every box to satisfy streamlined universal implementation with easy access, portability, turnkey utility, concise results, and the measurement integrity necessary for clinical assessment. But that's ok because 1) we’re still amidst these technological developments and it is possible we may have such tools in the future, 2) there are indeed some CVMC solutions ready for specific applications, and 3) while fast and easy use of a technology is appreciated, convenience should never come at the cost of measurement integrity.
In contrast to MBMC, CVMC is sometimes referred to as markerless motion capture (MLMC) due to the absence of sensors (reflective markers) affixed to the body. However, the term markerless encompasses a wide variety of systems built for an equally wide variety of applications and measurement performance standards. Consider the broad term automobile, which can be used to reference both a Jeep Wrangler and Formula 1 race car, even though these automobiles are designed for wholly different purposes with different performance standards to address different functional utilities. We would not expect the Jeep Wrangler to lap a professional racetrack with the same speed and agility as the Formula 1 race car, nor would we expect the Formula 1 race car to clear sand dunes as well as the Jeep Wrangler. Similarly, not all markerless systems quantify human movement with the same precision and accuracy. They are not intended to. Accordingly, when someone states they used markerless to measure human movement, prepare your follow-up questions. In the remainder of this section, we will outline some key features of CVMC systems commonly used to measure human movement to better arm the reader for engaging in the discussion. This is a high-level overview, so prudent readers are encouraged to explore the references provided for more detail.
Some pipeline infrastructures being built (e.g., MediaPipe (Lugaresi et al., 2019), Theia3D (Kanko et al., 2021), or OpenCap (Uhlrich et al., 2023)) enable users with limited computer coding experience to implement a specific computer-vision algorithm. This is precisely the type of work that needs to be done to bridge the gap between developers and end-users. However, “A common mistake that people make when trying to design something completely foolproof is to underestimate the ingenuity of complete fools.” - Ford Prefect, friendly alien researcher, finds himself in an unexpected predicament after jumping through a tall building window which it turns out was not rocket-proof from the inside and “deposited dizzily” on a 1-foot window ledge outside the 13th floor, and realizes that the way to solve this unfortunate problem is to think outside the box of expected uses engineers considered when designing that window ledge, and alas! It was outside the design scope to consider a person would ever BE perched on the window ledge, so simple window latches were used. He breaks back into the building by picking the window lock with a credit card and climbing back inside to safety, (Adams, 1979–1992; Book 5, Chapter 12).
To avoid disappointment, be aware that even with these pipelines in hand, CVMC systems are not turnkey. Some work and education are needed on the part of the end-user to: 1) ensure selection of the appropriate CVMC system for your application, 2) ensure the CVMC system is being implemented as intended, and 3) recognize whether measurement quality of the selected CVMC system is sufficient for the application. Most CVMC systems utilized to track human movement can be categorized by the camera quantity/type used to capture images and the algorithm method used for human pose estimation. Additionally, practical considerations of environment, movement task, and camera angle/distance to the individual will affect the quality of the measurement taken by the CVMC system.
Camera selections range from single (monocular) to multiple (x2, x4, x8+) views. The build of multi-camera CVMC systems is most like traditional MBMC systems, using overlapping views to triangulate the 3D position of a tracked body segment or key point. The benefit of multiple cameras is reduced occlusion, leaving more opportunities for direct measurement from the image. A consequential drawback to this approach is the fast accumulation of large data files. While these systems are portable for set up in various locations (McGuirk et al., 2022), which reduces the burden on studied individuals by bringing the measurement tool to them, some burden remains with the end-user to coordinate and implement the effort. Some may consider multi-camera systems as ‘just another lab solution’, unsuitable for clinical applications. We hold a contrasting view that the notably reduced burden on the studied individual, paired with the high integrity of the 3D measurement, makes it a strong contender for clinical assessment. The utility of a multi-camera CVMC system may perhaps be best appreciated where semi-permanent installation, like in a rehab gym or clinic-adjacent low-traffic hallway (Outerleys et al., 2025). Conversely, there are also CVMC systems using 1–2 cameras with benefits including the use of off-the-shelf hardware (e.g., smart phones, action/travel cameras) and ease of portability with fast setup and breakdown in multiple locations. Their drawback, however, is that only 1–2 views are captured for a movement performed in 3D space, limiting the information necessary for accurate, precise, reliable 3D measurement. Some cameras include depth sensors which extract 3D information from the 2D image. These cameras are appropriate to track small, single plane movements in real-time for telehealth applications (Knippenberg et al., 2017) which are conducted in smaller, light-controlled capture volumes (Scott et al., 2022). At present, CVMC systems with 1–2 camera views provide reasonable measurement for single plane movements in limited capture volumes (e.g., measurement of target accuracy during reach–to-point) but do not yet support assessment of 3D movements (Scott et al., 2022; Wade et al., 2022). The continued development of algorithms using 1–2 cameras may change this in the future (Needham et al., 2021). However, the current state of CVMC is such that when only 1–2 cameras are used, measurements are inadequate to provide 3D information required for clinical assessment especially in the presence of motor dysfunction where behavior is not predictable or isolated to a single plane (i.e., gait (Poomulna et al., 2025; Wang et al., 2024) or upper extremity activities (Kim et al., 2025)).
Presently, 3D finger movements such as pinching and grasping remain inaccessible to a streamlined CVMC solution due to inherent occlusion of the digits. Multiple camera viewpoints are necessary to track the spherical movement space. Fingers also require higher resolution images to capture the precision and accuracy of these finer movements. As a result, separate cameras must be set up with some positioned closer to the hand and dedicated only to tracking finger movements, while others are used to capture corresponding arm and trunk movements.
Common computer vision algorithm methods for human pose estimation include: 1) tracking human segments (i.e., visual hull, silhouette tracing) or 2) tracking key points/landmarks on the human body. Recent literature reviews provide a helpful baseline understanding for different human pose estimation methods and explanation of the training data sets (Ascenso et al., 2020; Chen et al., 2019; Desmarais et al., 2021; Mathis et al., 2020). Broadly speaking, visual hull/silhouette tracing methods perform best when excellent lighting and a sharp contrast are maintained between the studied individual and background environment, while key point estimation methods perform best when trained from image data sets labeled for biomechanics-grade applications. Some open-source algorithms were originally intended for entertainment applications (i.e., video games, animations, films). It is important to note that such algorithms may incorrectly fabricate the position of a key point when occluded and are trained from datasets with minimal key points (i.e., only 25 in OpenPose), tracked by people inexperienced with anatomical feature identification. While such solutions may be sufficient for performance standards in animation, they are not for high-precision biomechanics applications. Because these tools are open source, there is the potential to retrain the algorithms using different data sets, which is a future development to monitor (Cronin, 2021). Conversely, there are commercially available CVMC algorithms already designed specifically for the rigor of 3D biomechanical assessment. However, a criticism of these systems is the lack of transparency to verify measurement source, and the accessibility to a broad range of users. Measurement performance of these tools can continue to be interrogated and benchmarked, but their proprietary design means we are unlikely to know exactly how measurement was obtained.
“We demand rigidly defined areas of doubt and uncertainty!” - Vroomfondel, fictional HG2G philosopher (Adams, 1979–1992; Book 1, Chapter 25).
To accomplish adequately defined levels of certainty and validation, ideally, comparisons would be made between video fluoroscopy and CVMC systems to establish measurement precision and accuracy of the predicted joint centers during dynamic motion. Comparisons to MBMC are appropriate, but complicated as previously noted, due to marker-based measurements also involving error (i.e., skin/clothing artifact, incorrectly placed markers). However, the relevant functional movements for most clinical populations have already been quantified using MBMC methods with parameters like minimal detectable change linked to clinically relevant improvement/deterioration. In these cases, it would be helpful to establish any systematic differences between MBMC and CVMC (Hansen et al., 2024), as well as where CVMC may excel or is still producing insufficient measurement quality. To date, some concurrent comparisons between MBMC and CVMC systems report systematic offsets in specific joints (pelvic tilt, ankle). These are, in part, due to definitional differences in the biomechanical models used, not necessarily the measurement integrity of the raw data (Song et al., 2023; Wu et al., 2002). Some might describe these definitional differences between biomechanical model as “… so hip they have difficulty seeing over their own pelvis.” - Zaphod Beeblebrox, the solipsistic president of the HG2G Galaxy describing himself as cool (i.e., hip) (adapted from Adams, 1979–1992; Book 2, Chapter 6).
Also, the higher inter-trial variability reported in some, but not all, CVMC systems may impact clinically relevant measurements (Poomulna et al., 2025). For clinical populations (e.g., stroke, cerebral palsy) in which it is necessary to track movement in multiple planes, especially about the pelvis, hip, or ankle, substantial progress towards developing domain-specific training datasets and establishing relevant benchmarks is required to support clinical implementation and utilization of CVMC (Needham et al., 2021). Informative systematic reviews of CVMC for clinical application have been done by Wade and colleagues up through 2022 (Wade et al., 2022) and again in 2023 by Lam and colleagues (Lam et al., 2023).
To summarize, these are some elements which the motion capture community can implement to ensure CVMC remains mostly harmless to both the future integrity of motion analysis and clinical deployment:
Establish CVMC measurement benchmarks. When feasible, dynamic motions should be concurrently quantified and compared between CVMC and dynamic fluoroscopy. More realistically, this will be done by comparing concurrent measurements between CVMC and MBMC systems. Not all functional movements can be tracked by MBMC due to significant marker occlusion (i.e., a person walking with a walker and closely guarded by a physical therapist) though some CVMC algorithms still produce a human pose estimation. When direct measurement is not available for comparison, it remains an important consideration for the motion capture community to consider when the human pose estimation produced by the algorithm is sufficient. Different clinical populations, functional movements and situational changes (i.e., clothing, room lighting) should be interrogated to understand when CVMC will perform to standard for clinical use. Establishing measurement benchmarks should not be limited to a few experienced users. Rather, performance should be reported by multiple independent users of varied experience to validate the technology is indeed appropriate for clinical deployment. Standardize pairing reports of CVMC measurements with CVMC system specifications. Reported details should include both hardware (camera type/quantity/position/sampling frequency and software (algorithm type/version/access date) specifications. This strict pairing of measurement reports referenced to measurement tools was previously adhered to in the early days of MBMC, though that practice relaxed some with the universal improvements in MBMC measurement quality. CVMC measurement quality is still highly dependent on details of the tool and how it is used, so reported measurements should be contextualized by this information. Communicate with best practices and acceptable uses for specific CMVC systems. These should be updated as CVMC develops. While some of this knowledge comes from understanding how differences like lighting, camera position or an algorithm's training data set affect measurement quality, there are general considerations which could be better communicated within the motion capture community. For example: ‘CVMC algorithm A applied to data collected using camera quantity B and camera positions C can be used to quantify functional movement D of the knee in the sagittal plane, but is not yet adequate to measure pelvic tilt’. Transparent computer vision algorithm builds using domain-specific training data sets. CVMC systems using 1–2 cameras offer an appealing path to clinical translation due to their potential for better end-user accessibility and easier implementation. However, for these systems to have clinical utility, they must be able to capture full body human motion across multiple joints and planes with improved accuracy and precision. Realization of this goal requires algorithms which have been trained using expertly labeled, domain-specific datasets (Needham et al., 2021). Development of targeted and efficient post-processing pipelines. This item is more a consideration towards the successful implementation of CVMC in clinical practice, though the improvement would also benefit clinical research. Clinics are busy places, and the limited time practitioners can dedicate to each patient must be optimized. To ensure the utility of CVMC in the clinic, measurements need to be provided quickly, with minimal input required from the end-user. One example of this is an algorithm that can stay focused on tracking the patient in a populated room. Another example is easy to use software tools which allow the clinician to queue up the steps to automatically process a specific movement or clinical test and produce meaningful metrics/figures, requiring minimal input from the end-user. Development of such targeted pipelines will require partnership with clinicians to design analyses which address their clinical questions.
The Clinician's Guide to the Future
“Anything that happens, happens.
Anything that, in happening, causes something else to happen, causes something else to happen.
Anything that, in happening, causes itself to happen again, happens again.
It doesn’t necessarily do it in chronological order, though.”
- HG2G Narration (Adams, 1979–1992; Book 5, Preface)
Human motor systems neuroscience needs accurate measurement of behavior (Krakauer et al., 2017). Without it, we lack a clear understanding of the causal nature of disordered motor control (Chen & Patten, 2008). Equally important, accurate measurements afford a means to determine how (dare we say if?) rehabilitation interventions remediate motor function for stroke survivors.
With similar intent, the Stroke Recovery and Rehabilitation Roundtable (SRRR) established a game changing set of standards for stroke recovery research (Bernhardt et al., 2017). Their inaugural 2017 meeting generated a core set of consensus-based recommendations that argued a strong case for the addition of quantitative movement parameters (i.e., kinematics, kinetics) to improve design and outcomes for stroke recovery assessment (Kwakkel et al., 2017). While it is generally agreed that quantitative movement analysis is an important tool for assessing pathomechanics and can provide indicators of patient recovery with improved accuracy compared to current clinical outcomes, the SRRR stopped short of recommending specific measurement parameters. This is, in part, because motion capture technologies have not been widely available to clinicians (Buckley et al., 2019; Kwakkel et al., 2017).
In 2023, the SRRR updated these guidelines (Van Criekinge et al., 2023) with specific recommendations for use of quantitative motion analysis in independent ambulators (defined as having the ability to walk with or without aids, but without human assistance (Holden et al., 1986)). Progress? Perhaps. While more than 70% of survivors walk independently six months after stroke, this broad definition of “independent walking” encompasses a vastly heterogenous set of motor impairments and consequent gait patterns. The SRRR's compromise thus recognized pragmatics, acknowledging the limited feasibility towards implementing traditional MBMC in persons requiring hands-on or close-guard assistance, walking aids or orthotics. This incongruence between actual functional manifestations of walking in stroke survivors and the realities of acquiring usable motion capture data highlights the critical need for and significance of current developments with CVMC.
“What do you get when you multiply 6 by 9? 42. That's it. That's all there is.” - Arthur Dent, incorrectly interpreting his subconscious to identify the “Ultimate Question and Answer of Everything” using Scrabble pieces (Adams, 1979–1992; Book 2, Chapter 33).
It is notable that in the HG2G fictional world, with so much technological advancement (time travel, sentient robots, etc) their society remained unchanged, miserable even, while passively waiting for technology (supercomputer, Deep Thought) to give their lives meaning. In contrast, when the fictional HG2G Earthling, Arthur Dent is presented with this meaningless “Answer to the Ultimate Question of Life” (42!), he recognizes the absurdity of the situation and chooses to embrace that absurdism, instead of despairing in a meaningless existence. He subsequently creates his own future direction, rich with personal meaning. Regardless of your own philosophies about the meaning of life, this approach of taking personal responsibility to achieve the desired outcome is something most of us can appreciate.
We can take a similar perspective when considering what successful implementation of CVMC might look like in clinical practice. Otherwise, to echo Douglas Adams’ prescient concern, “we are stuck with technology when what we really want is stuff that just works” (Adams, 2002). Of itself, the presence and availability of CVMC in clinical spaces does not automatically advance clinical practice; it is merely the tool that opens the door to that potential. It is the clinician that advances clinical practice by taking personal responsibility to walk through that door and develop new, targeted stroke rehabilitation interventions informed by motion analysis as the SRRR envisions (Van Criekinge et al., 2023).
As clinicians, we often lament the need for better data to evaluate our patients and identify their actual impairments or changes in response to rehabilitation. To some degree, this lament is analogous to the proverbial dog chasing a car. What happens when the dog catches the car? The field is on the verge of enabling clinicians to ‘catch the car.’ But now what? Motion capture generates data and lots of it. The clinicians driving this paradigm shift in stroke rehabilitation will need to consider how they expect to interact with biomechanical data. Do they expect data presented as a video clip of a patient's biomechanical model, allowing the clinician to take in more information and multiple perspectives with each replay and different vantage point? Or perhaps they expect data plots depicting joint angle displacement while a patient completes a specific movement task, with potential to reveal subtle changes over time at follow-up appointments. Or is the expectation to characterize the quality of a patient's movement using metrics pulled from the biomechanical data? Revisiting the point that CVMC systems are not turnkey also means the data output can be tailored to answer your clinical question. Thus, there is not a single correct answer to the question “how will you use the data?” What aspects of the data are most useful to your clinical decision making? Would your intervention plan change once informed by these data? We hope so. Additionally, these data could be used to document progress and milestone attainment, helping to justify the necessity of current therapeutic services. In many cases, clinical therapists may already be delivering interventions as effectively and efficiently as possible. Given the frequent pressure to reduce or limit therapeutic services, CVMC could serve not only to enhance care but also as an objective means to validate and support existing clinical paradigms without adding undue burden. In the next section, we will include some examples of clinical questions that arise in stroke rehabilitation and how motion analysis may elucidate the outcome.
“A towel is about the most massively useful thing an interstellar hitchhiker can have. It has great practical value keeping the hitchhiker warm, dry and protected when faced with myriad situational challenges from the cold moons of Jaglan Beta to the noxious fumes of the Ravenous Bugblatter Beast of Traal. The towel also holds immense psychological value because it implies any traveler who can hitch the length and breadth of the galaxy, struggle against terrible odds, win through, and still hold onto their towel is clearly a person to be reckoned with.” - Text from The Guide (adapted from Adams, 1979–1992; Book 1, Chapter 3).
Relating this metaphor to implementing motion analysis in clinical practice, intrepid clinicians hitchhiking though the motion capture galaxy, should also carry a towel which in this case is their clinical question. In fact, Deep Thought had its own theory as to why society found its “Ultimate Answer” (42!) meaningless. When prompted to recheck its calculations, the supercomputer replied: “I checked it very thoroughly, and that is quite definitely the answer. I think the problem, to be quite honest with you, is that you’ve never actually known what the question is.” - Deep Thought (Adams, 1979–1992; Book 1, Chapter 28).
Similarly, the path to meaningful and effective clinical implementation of motion analysis starts with identifying the clinical question you want to answer. “The scope of clinical questions posed in stroke rehabilitation that could be supported by quantifying movement function is big. Really big. You just won’t believe how vastly hugely mindboggling big it is. I mean you may think it's a long way down the road to the chemist, but that's just peanuts to the clinical questions we can address using motion analysis.” - The Guide Introduction (adapted from Adams, 1979–1992; Book 1, Chapter 8).
Let's consider some common clinical scenarios that occur in stroke rehabilitation and how motion analysis could enhance clinical problem solving.
Clinical Scenario 1 - Balance Assessment. Current clinical assessments of posture and balance range from observation of postural sway and comparison between eyes open, eyes closed, potentially standing on an unstable surface, dual-task (e.g., serial 7's, counting, letter recognition, etc.), or sensory conflict conditions. Additionally, one may observe both internal and external perturbations, stepping, or gait initiation. Typically, one observes and grades posture and stability on an ordinal scale that may be linked to an instrument such as the Mini-BESTest or more typically may be clinician-specific. Another common approach is to time the patient's performance conditions, scoring them ‘within normal limits’ if they achieve 10 or 30 seconds without stepping or requiring hands on clinician support. Considerable nuanced detail eludes us while attending to patient safety and flicking one's stopwatch. Using the same amount of space, and for the same amount of clinician time and effort, CVMC can track kinematics of the body center of mass (COM) – providing means to appreciate this nuanced detail. Data processing pipelines can be developed to quickly calculate quantitative metrics, for example, COM area, path length, velocity, acceleration, variability, etc. Moreover, with a CVMC system available in the clinical setting, these measurements can be repeated easily over time and document the patient's performance, avoiding the all too familiar clinical experience of … 'they seem different this time' … 'I know they performed better today than a month ago' … Quantitative measurements of COM trajectory, velocity, etc. can distinguish important differences in specific conditions – sensory conflict, dual-task, movement initiation – which would aid the clinician in developing their treatment plan. A generalized program to improve postural control will certainly help most any patient, but an intervention specifically targeting a dual-task movement initiation condition might have a much more significant impact for a certain individual. For that individual, you’ll be a rehabilitation rockstar! While this brief example focuses on a single parameter, body COM, by using 3D motion analysis it is possible to acquire whole body measurements and evaluate multi-segmental coordination and responses. Here is where it is up to the clinician/end-user to develop the foundational knowledge to understand what parameters are most informative to their clinical interpretation.
Clinical Scenario 2 - Upper Extremity Motor Control. Clinical assessments of upper-extremity motor impairment and function including the: Fugl-Meyer Motor Assessment, Action Research Arm Test, or Wolf Motor Function Test involve observing the patient perform tasks of gross and fine motor function and grading them on an ordinal scale or timing task completion. As noted above with balance assessment, these clinical instruments fail to capture nuanced details of disordered motor control. Additionally, while psychometrically validated, some of these instruments involve a non-trivial ceiling and/or scores lack specificity. For example, three individuals each receiving 34/66 points on the UE Fugl-Meyer could reveal distinctly different impairment patterns. In this case, the clinical score is not helpful at all. In contrast, using a CVMC system to acquire UE kinematics during a simple reach-to-point or reach-to-grasp task, it is possible to acquire a large array of quantitative metrics ranging from: individual joint range of motion, end point error, trajectory path length, variability, and determine whether inter-segmental movement patterns are simply deficient (i.e., reduced ROM and velocity across all joints), a proximal joint compensating for distal limb impairments, or the pattern differs across repetitions indicating the patient has a severely compromised internal model. Each of these presentations would inform a different intervention. Again, patients can be evaluated without requirement for specialized clothing (or lack thereof) and repeated measures over time can be grouped to track progression.
Clinical Scenario 3 - Gait and Posture. Our group has conducted studies of posture, gait, and turning in clinic and community-facing spaces with people wearing their own clothing and footwear. There are myriad insights provided by full 3D kinematic data of individuals that extend beyond the well-recognized description of gait following stroke (i.e., slow speed, asymmetrical step length, reduced paretic limb support time, etc.) revealing other tractable intervention targets including traditionally musculoskeletal issues such as reduced thoraco-lumbar spine range of motion limiting weight shift onto the less affected side or kyphotic posture and forward head reinforced by likely cervico-thoracic spine displacements that contribute to limited step/stride length and, by extension, walking speed. An important opportunity afforded by quantitative 3D data is moving beyond the assumption that all stroke survivors reveal a similar set of gait deviations that result solely from neuropathology. Not only do kinematics characterize details of an individual's gait pathology, but these phenotypes also reveal whether the movement behaviors represent compensation – for either neural or biomechanical pathology – or a restorative pattern and capacity for recovery. The important distinction, and paradigm shift, is that rather than templated ‘one size fits all’ approaches based on heuristic therapeutic principles, interventions could be individually tailored to: 1) target specific functional and structural impairments, 2) promote recovery of neurologic function, 3) remediate compensations, or 4) in necessary cases accept that external devices and equipment are indeed necessary. Again, repeat studies can inform the rate and magnitude of progress on the trajectory of recovery and document the efficacy of rehabilitation interventions. An additional insight we have gained stems from cases where the patient is able to walk without an assistive device, but we observe their gait behavior is markedly disrupted in the presence of the assistive device. While safety is of paramount importance and we are not advocating to forego the assistive device, it opens a significant question regarding the challenge presented to learn/adopt a novel, biomechanically inefficient, and non-automatic walking pattern in the context of neuropathology.
In previous work, we illustrated how kinematic data acquired with CVMC can enhance clinical assessment and inform alternative, potentially more effective, interventions for persons affected with neuropathologies, including stroke (McGuirk et al., 2022). In addition to a tool for data-informed clinical decision making, breakthrough technologies such as CVMC offer an exciting opportunity to interrogate new hypotheses and develop genuinely effective rehabilitation approaches. By keeping your clinical question in hand, you can determine how motion analysis may contribute to a more meaningful clinical answer, as well as guide selection of the CVMC system specifications you use.
“ Application. Identify the movement task(s) of interest (i.e., gait/turns, balance/stability, sit-to-stand, reach/grasp), and movement plane(s) of interest (sagittal, frontal, coronal). Is the purpose to assess (measurement of motion) or interact (telehealth, real-time biofeedback)? What is the clinical population of interest and key characteristics of their movement dysfunction? For example, if the application is assessing compensatory upper extremity strategies of a stroke survivor during a reaching task, a monocular CVMC system would provide you with information in a single plane (sagittal or transverse, depending on camera position). However, to capture compensatory movement strategies in multiple directions at the same time, multiple cameras are needed at this time. Quantification. Identify metric(s) and/or data visualizations of interest. What are the standards of these measures for the clinical population of interest (e.g., range, normative values, minimal detectable change)? What specific time-points (e.g., terminal stance of the gait cycle) and features in movement are of interest (e.g., base of support while walking or total displacement of the body center of mass when comparing quiet standing between eyes open and eyes closed)? For example, if you wanted to quantify hip rotation during the swing phase of gait, pay attention to CVMC benchmark reports of measurements at that time in the gait cycle. While the average root mean square difference (RMSD) over the full gait cycle may fall below an acceptable RMSD value, differences during swing may still be too high. Monitoring updates in benchmark reports as these technologies continue to develop will help you determine when the measurement integrity of a specific CVMC system is sufficient to address your biomechanical metric of interest. Usage/Environment. How often (daily, weekly, occasionally) will the system be used and how many people will use it? Is your priority immediate access and fast turn-around of a testing assessment by way of a semi-permanent camera setup (e.g., wall mounted in a dedicated area); mobile access and the ability to test in multiple locations (e.g., tripods); or remote access (e.g., personal computer or smartphone). Identify the space/room necessary to capture the movement of interest. Beyond the practicality of measurement, implementation in active, real-world clinics is complicated. Space is at a premium everywhere, but especially in the clinic where rooms are shared and designed for multipurpose use. What is daily traffic and space occupancy? Do you need a private room with controlled access, or can you work in a more public space like a hallway or rehab gym? If the latter, how disruptive is clinic traffic to your measurement? If using a tripod in a public space, consider when and how often to calibrate camera positions due to accidental bumping. For example, a clinical assessment like the Timed-Up-And-Go (TUG) requires more floor space than just performing chair rises, so cameras would need to be positioned in such a way to capture the full maneuver. If you and your colleagues are performing the TUG in an active rehab gym, multiple wall mounted cameras would remain accessible at all times and wouldn’t be accidentally bumped – which would negate any data acquired. The algorithm for human pose estimation should be able to handle multiple people in the rehab gym, and the processing pipeline would only be applied to your patient. Alternatively, if you are a nomad clinician, performing the TUG in different community and home settings, you might want a monocular CVMC system which can be quickly set up using a single tripod, with the understanding that current monocular systems are limited to quantifying data accurately in one single plane/joint at a time. Depending on the quantifications of interest (see item 2 above), a monocular CVMC system may or may not be ready for this use case but continue to monitor benchmarking reports as this may change in the future. Data Storage/Security. Before you proceed, consider what IT/data security/IRB constraints need to be navigated. Does your institution allow/provide 3rd party cloud data storage of patient-identified videos, and does your patient consent to having their videos stored by a third party? If assessments are performed in public areas, how are incidental captures of images handled (e.g., people walking in background, people sitting in waiting areas, family members, etc.)? What resources are available for backing up video data files? Price. Consider costs to acquire, operate, and maintain the system. Is a powerful computer with a specific graphics card necessary, or will all processing be done remotely in the cloud? Note price includes not just fiscal considerations. What burden of time are you and your team willing/able to absorb in terms of: system setup, post-processing pipeline development, and time to process/manually correct misidentified data? Again, considering our example of a multi-camera CVMC system wall-mounted in a busy rehab gym, set up time is minimized to camera calibration, though longer post-processing times may be necessary to ensure only the patient's pose estimation is analyzed. A monocular camera system transported into the field can be done by a single person with less time and effort than a multi-camera system, but that value must still be weighed against the limited data that can be used for clinical practice at this time.
As you can see, selecting the appropriate CVMC system really depends on your clinical question. By keeping your clinical question in mind and staying informed of the next CVMC technological advancements, you may confidently navigate this new and changing landscape. Expect ongoing development in the CVMC system specifications within these four topics:
Camera Views/Count. How many camera views are needed to quantify your movement(s) of interest with the precision and accuracy required for your specified clinical population(s)? Algorithm Type/Training. What method of body pose estimation is used (visual hull/silhouette tracing or tracking key points with deep machine learning)? How was the algorithm trained to handle body segment/key point occlusion? Benchmark/Validation. What peer-reviewed work is available specific to your selected: a) CVMC system (camera count and algorithm type), b) human movement, and c) clinical population? Note details such as sampling frequency, camera view/angle, and version/release date for the trained algorithm. Have multiple validations from independent groups produced similar reports? Is the reported measurement quality only achievable by a practiced team, experienced with (or having developed) the CVMC system, or is it reasonable to expect a new user to achieve similar performance? If you are interested in a specific joint and/or time point during a movement (i.e., hip rotation at terminal stance), how does the CVMC system perform on those parameters, because that is what matters to you now. Technical Skills. How accessible is this system to you? Is it an open-source algorithm intended to be built into your own code, or is a processing pipeline also provided? What are the hardware set-up and camera calibration steps? What post-processing steps occur between capturing a movement on video and producing your desired metric or plotted figure? Technical skill is important, but last on this list, because there are friendly engineers (alien researchers) familiar with biomechanics and motion capture to meet along the way and be your hitchhiking companion.
“ “At some point, the storm will definitely abate, and what thunder there is now will grumble over more distant hills, like a man saying ‘and another thing …’ twenty minutes after admitting he's lost the argument.” - HG2G Narration (adapted from Adams, 1979–1992; Book 4, Chapter 3).
But if we move too quickly and overpromise or misrepresent the current capabilities of this technology, we risk compromising its potential for both clinical research and clinical practice and are likely to end up with measurements “almost, but not quite, entirely unlike human movement” - Arthur Dent, remarking on a particularly unfortunate cup of tea made by a Nutri-Matic Machine (adapted from Adams, 1979–1992; Book 1, Chapter 17).
In parallel to the continued developments of this technology, the arrival of CVMC opens the door to a potential paradigm shift towards more targeted stroke rehabilitation methods, should clinicians choose to walk through and embrace the information offered by data acquired from motion analysis. We hope curious clinicians will be emboldened to grab their towel and join this exploratory journey. After all, it is just the quantification of human movement. It's not like we asked CVMC to make a cup of tea.
Footnotes
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.