Towards a mobility diagnostic tool: tracking rollator users' leg pose with a monocular vision system

Area of Science:

• Biomechanics
• Computer Vision
• Robotics

Background:

• Cognitive assistance for rollator users may impair stability by reducing attentional capacity.
• Accurate tracking of rollator user pose is essential before implementing cognitive assistance systems.
• Existing markerless vision systems primarily focus on sagittal plane motion, not coronal plane dynamics.

Purpose of the Study:

• To estimate the 3D pose of lower limbs for rollator users from monocular image sequences in the coronal plane.
• To address the challenges of a single frontal view and subtle coronal plane motion.
• To evaluate the effectiveness of different appearance models (cues) for pose estimation.

Main Methods:

• Utilizing a Bayesian probabilistic framework to estimate leg limb pose.
• Exploring multiple visual cues within the probabilistic model.
• Focusing on evaluating the performance of the appearance model, specifically texture and color cues.
• Conducting preliminary experiments with manually initialized appearance offline.

Main Results:

• Appearance models conditioned on the specific visual characteristics of a rollator user demonstrated superior performance.
• Texture and color cues outperformed more general appearance cues in estimating lower limb pose.
• The approach requires manual offline initialization of the appearance model.

Conclusions:

• Customized appearance models are effective for estimating rollator user lower limb pose in the coronal plane.
• Texture and color are valuable cues for this specific application.
• Further research may focus on automating appearance model initialization for enhanced usability.