A video-based model preserved the broad shape of several human movements, but it was less dependable on exact position and fine coordination, according to a new preprint. In a case-based evaluation, the system reconstructed marker paths during assisted sit-to-stand movement, projected knee and trunk descriptors, hip-knee coupled motion and hopping. The results suggest a feasibility foundation, not evidence that the same performance would hold across participants.
The evaluation covered assisted pediatric movement, unassisted pediatric movement and multidirectional hopping. The human cases included children with cerebral palsy, typically developing children and one young adult. Each displayed test condition was represented by one participant, so repeated movements measured within-case performance rather than between-participant generalizability.
Inside the test
The researchers compared synchronized video with reference marker trajectories. Twelve Vicon cameras recorded the reference at 100 Hz, while two FLIR cameras supplied video at 50 Hz. The target was a marker path, not a direct estimate of an anatomical joint centre. The pipeline used view-specific rank-reduction autoencoders, which compress each camera view into a smaller internal representation, followed by a feedforward multilayer perceptron to predict marker coordinates. Temporal smoothing was applied after the prediction. For testing, one recording per condition was held out, while the remaining sequence data were split into 80% for training and 20% for validation.
Position was uneven
In the assisted cerebral-palsy case, the right-back marker was tracked most closely along the Z axis, with a mean absolute error of 17.76 mm across four test repetitions. The study used mean absolute error, which records the average size of prediction differences, as its main discrepancy measure. The X-axis error was 27.00 mm, while the Y axis reached 59.47 mm. R², a complementary score for trajectory fit, was 0.835 on Z, 0.564 on X and -0.164 on Y. The model preserved some overall progression on Y, even though its positional agreement there was weak.
Other descriptors showed why a single score did not tell the whole story. Across four test repetitions, projected knee descriptors had mean absolute errors of 5.65 degrees in the sagittal plane, 4.46 degrees in the frontal plane and 14.55 degrees in the transverse plane. The corresponding R² values were 0.944, 0.425 and 0.872. The sagittal descriptor had the strongest trajectory fit, while the transverse plane had the largest absolute error. For pelvis-relative trunk descriptors, the sagittal error was 5.38 degrees, compared with 4.10 degrees frontally and 3.45 degrees transversely. Yet R² was strongest sagittally at 0.818 and negative in the frontal and transverse planes, at -0.659 and -1.280.
When two joints move together
The model faced a harder test when the relationship between hip and knee movement was considered together. A cyclogram is a plot that pairs two movement measures, and in the cerebral-palsy case it showed recognizable coupled-motion features but differences in excursion and local geometry. Hip and knee flexion-proxy mean absolute errors were 6.16 and 3.38 degrees, while the paired-angle discrepancy was 7.60 degrees. Hip and knee R² values were 0.422 and 0.423. Numerical comparison was possible for 80.4% of normalized samples; 19.6% had no paired reference value.
In the typically developing case, the broad reference trajectory was more closely reproduced across three test repetitions. Hip and knee R² values were 0.879 and 0.960, while mean absolute errors were 7.17 and 4.67 degrees. The paired-angle discrepancy was 9.41 degrees, and paired data were available for 100% of normalized samples. The comparison shows that following the broad path of a coupled movement does not necessarily mean matching every detail of its excursion.
Hopping offered a clearer timing signal
The hopping test also reproduced the main timing signal. All 16 jumps were valid. Reference flight time averaged 0.234 seconds, compared with 0.242 seconds for the prediction. The bias was 0.009 seconds, the mean absolute error was 0.054 seconds and the root mean square error was 0.078 seconds. The reconstructed left-knee Z trajectory had a mean absolute error of 20.96 mm, with R² of 0.760 and Pearson correlation of 0.878. The dominant flight-time and displacement profiles were retained, although individual hop amplitudes differed.
A case study, not a population result
The study was a case-based feasibility evaluation, and its repeated-movement analyses addressed within-case performance rather than between-participant generalizability. One participant represented each displayed condition, and the repetitions were nested within those cases. The analysis was descriptive and included no population-level hypothesis tests, so it cannot estimate how accurately the framework would perform across a wider group.
The agreement figures also have a built-in boundary: intervals without paired reference data could not be independently checked. In the cerebral-palsy cyclogram, 19.6% of normalized samples fell outside numerical comparison. The reported results therefore describe the portions of the movement for which reference and predicted samples could be paired.
What remains controlled
De-identified data were not openly released because consent and institutional restrictions did not allow unrestricted sharing, although they may be provided on request under institutional and ethical conditions. Custom analysis code was available from the corresponding author during peer review and was planned for permanent repository release after publication. Cerebral-palsy data collection received support from the NIH and the Foundation for Physical Therapy Research, while computational development and secondary analyses had no dedicated funding. The authors report no competing interests.
Paper data and sources
Original title: Human movement reconstruction through latent structural representations during assisted transitions and multidirectional hopping
Authors: Guilherme Manna Cesar, Ligia Yumi Mochida, Chady Ghnatios
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-10-07
DOI: Not available
Original paper · Full text