Background Musculoskeletal diseases, such as Juvenile Idiopathic Arthritis (JIA), can lead to altered joint mobility and changes in gait pattern. While marker-based motion capture systems are considered the gold standard, their use is limited by high costs, technical complexity, time-intensive procedures, and specialized staff requirements. This study proposes and evaluates the feasibility of a markerless gait analysis approach for children of all ages, based on a machine learning pipeline, validates it against marker-based systems, and describes JIA-related gait differences. Methods Sagittal-plane gait videos were recorded using a standardized setup. A Mask R-CNN-based detector generated a bounding box around the body in each video frame. Within this box, a TCFormer model identified positions of predefined anatomical landmarks. Joint angles were computed from these landmarks on a frame-by-frame basis and assembled into continuous gait cycle trajectories. For validation, a subset of 12 participants was recorded simultaneously using a marker-based motion capture system and conventional video-recording. To ensure direct comparability between the two methods, physical markers in video-recordings were automatically detected and removed using a deep learning-based object detection and inpainting approach. Statistical comparisons were performed using Statistical Parametric Mapping. The estimated kinematic data were compared between JIA patients with lower-extremity involvement and typically developing (TD) peers, stratified by age and disease activity. Results Videos from 108 participants (60 JIA; 48 TD; age 1.8–17.4 years) were included. There was robust agreement between joint movement patterns of the markerless and the marker-based approach. Across all age groups, JIA patients with active arthritis had a reduced knee range of motion compared to TD (46.3<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(^{\circ}\)</EquationSource> </InlineEquation> vs. 49.8<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(^{\circ}\)</EquationSource> </InlineEquation>). Overall, differences in joint angle trajectories between groups remained limited. Conclusions The proposed single-camera, markerless pipeline can provide reliable and valid gait data using standard video equipment in children. Differences in gait patterns between JIA patients and TD peers were generally small, likely due to effective disease control. This method is adaptable across all ages, may facilitate enhanced gait monitoring, increase knowledge of functional abnormalities in JIA, and support therapeutic decisions. In the long run, it may also reduce costs by lowering infrastructure and setup requirements.

Markerless gait analysis for children and adolescents with juvenile idiopathic arthritis using a machine learning pipeline

期刊论文
出版模式：

开放获取
发表日期：

2026年4月28日

查看全文

Aseem Behl,
Dominik Papies,
Winfried Ilg,
Mareen Kraft,
Andrea Bevot,
Sandra Hansmann

摘要

Background

Musculoskeletal diseases, such as Juvenile Idiopathic Arthritis (JIA), can lead to altered joint mobility and changes in gait pattern. While marker-based motion capture systems are considered the gold standard, their use is limited by high costs, technical complexity, time-intensive procedures, and specialized staff requirements. This study proposes and evaluates the feasibility of a markerless gait analysis approach for children of all ages, based on a machine learning pipeline, validates it against marker-based systems, and describes JIA-related gait differences.

Methods

Sagittal-plane gait videos were recorded using a standardized setup. A Mask R-CNN-based detector generated a bounding box around the body in each video frame. Within this box, a TCFormer model identified positions of predefined anatomical landmarks. Joint angles were computed from these landmarks on a frame-by-frame basis and assembled into continuous gait cycle trajectories. For validation, a subset of 12 participants was recorded simultaneously using a marker-based motion capture system and conventional video-recording. To ensure direct comparability between the two methods, physical markers in video-recordings were automatically detected and removed using a deep learning-based object detection and inpainting approach. Statistical comparisons were performed using Statistical Parametric Mapping. The estimated kinematic data were compared between JIA patients with lower-extremity involvement and typically developing (TD) peers, stratified by age and disease activity.

Results

Videos from 108 participants (60 JIA; 48 TD; age 1.8–17.4 years) were included. There was robust agreement between joint movement patterns of the markerless and the marker-based approach. Across all age groups, JIA patients with active arthritis had a reduced knee range of motion compared to TD (46.3 \(^{\circ}\) vs. 49.8 \(^{\circ}\) ). Overall, differences in joint angle trajectories between groups remained limited.

Conclusions

The proposed single-camera, markerless pipeline can provide reliable and valid gait data using standard video equipment in children. Differences in gait patterns between JIA patients and TD peers were generally small, likely due to effective disease control. This method is adaptable across all ages, may facilitate enhanced gait monitoring, increase knowledge of functional abnormalities in JIA, and support therapeutic decisions. In the long run, it may also reduce costs by lowering infrastructure and setup requirements.