Clinical gait assessment normally requires a marker-based motion-capture laboratory: infrared cameras, reflective markers, strict setup standards and trained personnel, all of which put it out of reach of most clinics. This Idea covers a cheap alternative that predicts the clinically used gait numbers directly from one ordinary sagittal-plane video. OpenPose extracts 2D skeletal keypoints, and instead of feeding the resulting time series to a sequence model, the 124-frame clip is rendered as two 128x128 grayscale images: one writes each frame's 58 coordinate values as a row of a matrix that is then tiled to a square, the other plots the hip, knee and ankle landmarks of all frames on a single canvas. Two identical three-layer convolutional branches (32, 64, 128 filters, ReLU, 2x2 max pooling) encode the two images, their flattened outputs are concatenated to 65,536 features, and four fully connected layers with dropout regress a single scalar - the Gait Deviation Index, maximum knee flexion, step length or cadence, per side. Trained on 2,212 videos of 1,138 cerebral-palsy patients with a patient-disjoint 8:1:1 split, it beats a 1D-CNN and a spatial-temporal transformer by 10.6 percent on GDI and 12.0 percent on knee flexion while being smaller and faster, but is worse than both on cadence because nothing in the architecture detects gait cycles or strides, which is precisely what cadence counts. A student should be able to explain why turning time into an image axis works for whole-clip averages and fails for cyclic quantities, and to notice that the paper's own methodology section describes no attention layer despite calling the model a Transformer.
A short applied computer-vision research paper, accepted at IEEE ISBI 2025, that estimates clinical gait parameters from ordinary single-camera video instead of a marker-based motion-capture laborato…