You Only Look Once Version 7-Assisted Proportional-Integral-Derivative Trajectory Tracking in a Lower-Limb Rehabilitation Robot

This study presents a reproducible You Only Look Once version 7 (YOLOv7)-assisted proportional-integral-derivative (PID) method that leaves the low-level PID controller unchanged and uses sagittal red-green-blue (RGB) vision only to compensate reference-trajectory phase. Five lateral markers at the hip, knee, ankle, heel, and toe are detected, filtered, and converted into a four-state gait estimate and a bounded continuous phase correction.

Twenty healthy adults completed treadmill trials at 0.20 m/s, 0.30 m/s, and 0.40 m/s under normal visibility and partial sagittal-view obstruction. The locked detector achieved mean average precision at an intersection-over-union threshold of 0.5 (mAP@0.5) of 0.924 ± 0.021, 98.0% precision, 96.2% recall, and 12.7 ± 1.7 ms frame-level inference latency. Relative to PID-only control, YOLOv7-assisted PID reduced mixed-effects-estimated hip and knee root mean square error (RMSE) by 0.58° (95% confidence interval [CI], 0.53-0.63°) and 0.57° (95% CI, 0.51-0.63°), respectively; settling time decreased by 109.7 ms and torque-command root mean square (RMS) decreased by 0.69 N·m.