Abstract
This study presents a system for estimating the upper-body posture of a bicycle rider using only onboard sensors, without body-worn devices. Strain gauges on the handlebar and IMUs on the handlebar and frame serve as inputs to a LightGBM-based model estimating five joint angles: trunk lean, left/right shoulder, and left/right elbow angles. LightGBM was adopted to quantitatively evaluate sensor contributions via feature importance analysis. The rider’s upper body is modeled as a seven-link structure with independently estimated left and right limbs, enabling representation of asymmetric cornering motion that conventional symmetric models cannot express. Validation with one male participant across straight, curved, and figure-eight riding yielded a trunk lean RMSE of 9.97°, meeting the target of within 10°. Shoulder and elbow angles showed RMSEs of 18-24°, attributed to limited training data and motion capture errors from outdoor vibration. Feature importance analysis revealed that frame IMU features contributed 60–70% to trunk and shoulder estimation, while strain gauge features dominated elbow estimation at ~50%, suggesting handlebar load more directly influences elbow motion. Cornering trials confirmed an average left-right shoulder symmetry of 16°, demonstrating the advantage of the proposed model over symmetric approaches.
