JPJi Hoon Park

Ji Hoon Park

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Ji Hoon Park has a total of 2 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Real-Time Recovery of Missing Markers in an Optical Tracking Cluster Using IMU Signals and Rigid-Body Kinematics

    by Ji Hoon Park, Chang June Lee, Jung Keun Lee

    Optical motion capture (OMC) systems that use optical tracking clusters are widely used in rehabilitation and related clinical applications because they can accurately compute segment orientation. OMC-derived orientation is widely used as a reference for validating inertial measurement unit (IMU)-based orientation estimation algorithms. However, missing markers owing to occlusion occur frequently, which makes the OMC-based computation of segment orientation infeasible. Therefore, reliable missing marker recovery is essential in practice. Existing methods rely either on post-processing approaches that cannot be applied in real time or on IMU-based orientation of which errors can affect recovery performance. This paper proposes a real-time method for the recovery of missing markers using IMU signals and rigid-body kinematics but without IMU-based orientation. We implemented three methods based on rigid-body kinematics to evaluate recovery performance, all of which utilize intermarker distance: Method 1 (M1) using accelerometer signals; Method 2 (M2) using gyroscope signals; and the proposed Method 3 (M3) using both accelerometer and gyroscope signals. In addition, a Kalman filter-based method using IMU-based orientation (Method 4, M4) was implemented for comparison. Recovery performance was evaluated based on the number of missing markers (1 or 2) and the duration of the missing interval (5 s or 30 min). In the comparison among methods based on rigid-body kinematics, M1 showed degraded performance when two markers were missing, whereas M2 exhibited drift errors during prolonged missing intervals (30 min). In contrast, the proposed method (M3) consistently achieved low root mean squared errors of 2.25–4.23 mm and 4.05–7.08 mm for one and two missing markers, respectively, during the 5 s interval, and 1.10–3.52 mm and 3.72–5.92 mm during the 30 min interval. In comparison with the method using IMU-based orientation, M3 outperformed M4 under disturbed conditions. In particular, because M3 uses only accelerometer and gyroscope signals without relying on a magnetometer, it is inherently free from error propagation from magnetic disturbances to orientation estimation and, consequently, to recovery performance. In contrast, M4 showed degraded recovery performance due to disturbance-induced orientation estimation errors. These experimental results indicate that the proposed method enables the real-time recovery of missing marker data with high accuracy, even in disturbed conditions and over long missing intervals.

  • Open Access

    Effects of Gyroscope Bias Augmentation on Recurrent Neural Network Model Performance for Attitude Estimation

    by Ji Hoon Park, Chang June Lee, Jung Keun Lee

    The 6-axis inertial measurement unit-based attitude estimation is widely used in fields such as biomechanics and unmanned aerial vehicles. The 6-axis IMU signals, comprising 3-axis gyroscope and accelerometer data, provide crucial information for attitude estimation. However, measurement errors, particularly gyroscope bias, can reduce the estimation accuracy. Gyroscope bias augmentation can address this issue in end-to-end machine learning models; however, its effect on model performance has not been thoroughly investigated. This study investigated the effect of gyroscope bias augmentation range on the performance of an end-to-endmodel for attitude estimation, specifically focusing on varying biasmagnitudes. To this end, models were trained using data with different bias augmentation ranges. The performance of each model was then observed on signals with different bias magnitudes. The analysis revealed a generalization–precision trade-off relationship in the influence of bias augmentation range on the estimation performance. A narrower augmentation range exposed the model to smaller bias magnitudes during training, leading to superior performance within that range. However, the performance significantly degraded when encountering larger bias magnitudes that were not experienced during training. Conversely, models trained with a wider augmentation range were exposed to a broader spectrum of bias magnitudes, thereby enabling them to maintain relatively stable estimation performance even under larger biases. However, their performance deteriorated in the small bias range. These results indicate that, given known bias characteristics, selecting an appropriate bias augmentation range during training can help balance generalization and precision.