Deep Learning and Computer Vision Approaches for Markerless Human Gait Analysis Across Clinical Populations: A Narrative Review

Background: Markerless gait analysis based on deep learning (DL) and computer vision (CV) has advanced rapidly as an alternative to marker-based optical motion capture, which remains constrained by cost, laboratory infrastructure, and setup time. Method: This narrative review synthesizes eleven recent (۲۰۲۳–۲۰۲۶) primary studies validating DL/CV based markerless gait analysis systems against marker-based or instrumented reference standards, supplemented by six recent systematic/comprehensive reviews (۲۰۲۴–۲۰۲۶) used only as background context; no formal PRISMA search protocol or risk-of-bias scoring was applied. Results: Across diverse clinical and technical contexts including rehabilitation cohorts, pediatric rheumatology, post-surgical orthopedic recovery, sarcopenia and Parkinson's disease screening, and healthy-adult benchmarking markerless DL/CV systems consistently reproduced spatiotemporal gait parameters (step length, cadence, step time) with good to excellent agreement against marker-based references, with concordance or correlation coefficients frequently exceeding ۰.۸, although agreement was consistently weaker at the ankle joint and for gait symmetry measures. Joint-angle kinematics were more variable overall, with larger errors in the frontal and transverse planes and at the ankle and hip, and accuracy was sensitive to camera viewing direction. Task-specific fine-tuning of pose-estimation networks and camera inertial sensor fusion measurably reduced error relative to generic pretrained models. Conclusion: For several applications, DL/CV based markerless gait analysis is already close to clinically usable. What's still missing is standardized biomechanical modelling, validation in larger multi-site cohorts, and closer attention to camera placement before it can enter routine clinical practice.

کلیدواژه ها:

markerless gait analysis ، deep learning ، pose estimation ، clinical validation ، computer vision