Vision-based vs. IMU-based upper-limb pose estimation in assisted dressing: a comparative study of positional...

Area of Science:

• Robotics and Human-Computer Interaction
• Biomedical Engineering and Rehabilitation Technology
• Computer Vision and Machine Learning

Background:

• Accurate upper-limb kinematics are crucial for rehabilitation and assistive robotics.
• Real-world scenarios with occlusion and physical interaction pose challenges for current pose estimation methods.
• Vision-based methods show promise but their reliability for joint kinematics in complex tasks is uncertain.

Purpose of the Study:

• To systematically compare vision-based and wearable sensing approaches for upper-limb pose estimation during assisted dressing.
• To evaluate the performance of a monocular RGB-based Convolutional Neural Network (CNN) and an IMU-based method against a VICON motion capture reference.
• To assess both positional accuracy and kinematic agreement under real-world conditions.

Main Methods:

• Implemented a monocular RGB-based CNN and a temporally smoothed CNN variant.
• Utilized a wearable IMU-based reconstruction method.
• Compared both approaches against an inverse kinematics (IK) reference from VICON motion capture data, using participant-specific models.

Main Results:

• IMU-based estimation demonstrated consistent accuracy and stability in joint-angle reconstruction (mean absolute error ~12°).
• Vision-based methods achieved good positional accuracy (MPJPE ~0.20 m) but significant joint-angle errors (>80°), particularly with occlusion.
• Temporal smoothing improved positional consistency but not kinematic accuracy.

Conclusions:

• Current vision-based pose estimation methods have fundamental limitations for tasks requiring precise joint kinematics, especially with occlusion.
• IMU-based sensing offers a more reliable solution for accurate upper-limb kinematic estimation in assisted scenarios.
• Future reliable pose estimation may require integrating inertial sensing or biomechanical constraints.