Real-time human pose estimation on a smart walker using convolutional neural networks

Rehabilitation is important to improve quality of life for mobility-impaired patients. Smart walkers are a commonly used solution that should embed automatic and objective tools data-driven human-in-the-loop control monitoring. However, present solutions focus on extracting few specific metrics from dedicated sensors with no unified full-body approach. We investigate general, real-time, pose estimation framework based two RGB+D camera streams non-overlapping views mounted smart walker equipment in rehabilitation. Human keypoint performed using two-stage neural network framework. The 2D-Stage implements detection module locates body keypoints the 2D image frames. 3D-Stage regression lifts relates detected both cameras 3D space relative walker. Model predictions low-pass filtered temporal consistency. A custom acquisition method was obtain dataset, 14 healthy subjects, training evaluating proposed offline, which then deployed real equipment. An overall error 3.73 pixels 44.05mm were reported, an inference time 26.6ms when constrained hardware novel approach patient monitoring context walkers. It able extract complete compact representation real-time inexpensive sensors, serving as common base downstream extraction solutions, Human-Robot interaction applications. Despite promising results, more data be collected users impairments, assess its performance rehabilitation tool real-world scenarios.