Deep learning-based improvement of joint angle accuracy for uncalibrated multi-view markerless three-dimensional motion capture.

This study aimed to improve and validate the accuracy of joint angles estimated using a mixture of uncalibrated cameras (MUC), an uncalibrated multi-view markerless three-dimensional motion capture system, through bidirectional long short-term memory (BiLSTM) based correction. Twelve male participants performed bricklaying, assembly, and parcel loading while marker-based and multi-view markerless motion capture data were recorded simultaneously. MUC-derived joint angles were corrected using a BiLSTM model trained against marker-based reference values. Representative RMSE reductions were 48.9° to 25.0° for shoulder abduction-adduction in bricklaying, 11.3° to 5.8° for elbow flexion in assembly, and RMSE from 40.0° to 16.7° for knee flexion in parcel loading. Pearson's correlation and ICC improved for several motions, while Bland-Altman analysis showed reduced fixed and proportional bias and narrower limits of agreement in selected motions. These findings support BiLSTM-based correction for industrial ergonomic assessment.