GaitMirror: Design and Implementation of an Artificial Intelligence B ased Gait Recognition System for Evaluating the Effectiveness of Lower Limb Rehabilitation

Patients with lower limb instability-related diseases (e.g., stroke) have high fall risks, highlighting the need for effective rehabilitation and objective evaluation methods. Clinicians typically observe gait characteristics (e.g., stride lengths, gait stability, and coordination) using evaluation tools such as the Brunnstrom stages (BSs) to compare pre- and postrehabilitation statuses. However, this strategy heavily relies on clinical expertise and is susceptible to individual biases, posing challenges during objective evaluations. This paper proposes an artificial intelligence (AI)-based gait recognition system called GaitMirror, which is composed of an imaging camera, an edge AI computing device, and a touch screen panel. GaitMirror utilizes deep learning-based human pose estimation to detect keypoints in the captured images. By tracking the coordinate changes exhibited by specific keypoints during walking and applying an innovative subpixel-precision SBR algorithm, it computes multiple walking quantitative metrics (WQMs)—including length of a step (LS), number of steps (NS), walking speed (WS), step time variability (STV), and body sway angle (BSA)—to provide clinicians with objective gait data. Experimental results demonstrate that GaitMirror achieves sub-centimeter accuracy in LS measurement, thereby enhancing the precision of NS, WS, and STV calculations; moreover, its BSA measurements accurately reflect the body’s center-of-gravity shifts in both horizontal and vertical directions. In clinical trials, GaitMirror analyzed the gait characteristics of 20 stroke patients at different BSs, yielding consistent WQMs across patients within the same BS. These findings underscore the potential of GaitMirror to serve as a valuable complementary tool for clinical evaluation purposes.