Portable in-clinic video-based gait analysis: validation study on prosthetic users

Abstract

Despite the common focus of gait in rehabilitation, there are few tools that allow quantitatively characterizing gait in the clinic. We recently described an algorithm, trained on a large dataset from our clinical gait analysis laboratory, which produces accurate cycle-by-cycle estimates of spatiotemporal gait parameters including step timing and walking velocity. Here, we demonstrate this system generalizes well to clinical care with a validation study on prosthetic users seen in therapy and outpatient clinic. Specifically, estimated walking velocity was similar to annotated 10-meter walking velocities, and cadence and foot contact times closely mirrored our wearable sensor measurements. Additionally, we found that a 2D keypoint detector pretrained on largely able-bodied individuals struggles to localize prosthetic joints, particularly for those individuals with more proximal or bilateral amputations, but it is possible to train a prosthetic-specific joint detector. Further work is required to validate the other outputs from our algorithm including sagittal plane joint angles and step length. Code and trained weights will be released upon publication.

Clinical gait analysis using video-based pose estimation: multiple perspectives, clinical populations, and measuring change

“…Interest in video-based, markerless gait analysis has accelerated rapidly. Previous studies have used various approaches to move quantitative clinical gait analysis outside of the laboratory or research center and directly into the home or clinic 5,6,[13][14][15]17 . Here, we aimed to develop a single approach that addressed several outstanding needs, including the needs to accommodate multiple different types of environments/viewing perspectives, use of datasets in multiple clinical populations with gait impairment, measurement of both spatiotemporal gait parameters and lower extremity two-dimensional kinematics, and measurement of withinparticipant changes in gait.…”

“…Furthermore, there is a need for validation in additional adult and pediatric clinical populations, as previous work has shown that existing pose estimation algorithms have difficulty with tracking patient populations with anatomical structures that likely differ significantly from the images used to train the algorithms 13 . Thirteen of the participants with stroke used a cane; we did not observe instances where OpenPose mistakenly identified the cane as a limb.…”

“…We have previously used OpenPose 8 -a freely available pose estimation algorithm -to develop and test a comprehensive video-based gait analysis workflow, demonstrating the ability to measure a variety of spatiotemporal gait parameters and lower-limb joint kinematics from only short (<10 seconds) sagittal (side view) videos of individuals without gait impairment 11 . Others have also used a variety of approaches to combine pose estimation outputs and neural networks to estimate different aspects of mobility 5,[12][13][14][15][16] . This foundational work in using pose estimation for video-based gait analysis has demonstrated significant potential of this emerging technology.…”

Improved Trajectory Reconstruction for Markerless Pose Estimation

“…We note that the robust triangulation approach was described in a paper on multiview self-supervised learning (SSL) [27], where the information between views is used to improve the geometric consistency between 2D keypoint detectors, and we plan to implement this for fine-tuning keypoint detectors. This would have the added benefit of allowing learning from a diverse clinical population, which can address limitations we have previously noted such as tracking the location of prosthetic limbs [26]. Keypoint detectors would also be improved for biomechanics by learning denser keypoints over the trunk, as their absence can limit understanding pelvis movement and reconstructing hip angles without training additional models to mitigate this [3].…”

“…However, the DLT is sensitive to outliers, which will occur if another person occludes the view and has their joints detected, or if a joint is misdetected in one view, such as for people with limb differences like prosthetic users [26]. A common approach to outlier rejection in 3D reconstruction is RANSAC, but this is a slow algorithm.…”