Computer Vision-Based Hand Tracking and 3D Reconstruction as a Human-Computer Input Modality with Clinical Application
Abstract
The recent pandemic has impeded patients with hand injuries from connecting in person with their therapists. To address this challenge and improve hand telerehabilitation, we propose two computer vision-based technologies, photogrammetry and augmented reality as alternative and affordable solutions for visualization and remote monitoring of hand trauma without costly equipment. In this thesis, we extend the application of 3D rendering and virtual reality-based user interface to hand therapy. We compare the performance of four popular photogrammetry software in reconstructing a 3D model of a synthetic human hand from videos captured through a smartphone. The visual quality, reconstruction time and geometric accuracy of output model meshes are compared. Reality Capture produces the best result, with output mesh having the least error of 1mm and a total reconstruction time of 15 minutes. We developed an augmented reality app using MediaPipe algorithms that extract hand key points, finger joint coordinates and angles in real-time from hand images or live stream media. We conducted a study to investigate its input variability and validity as a reliable tool for remote assessment of finger range of motion. The intraclass correlation coefficient between DIGITS and in-person measurement obtained is 0.767- 0.81 for finger extension and 0.958–0.857 for finger flexion. Finally, we develop and surveyed the usability of a mobile application that collects patient data medical history, self-reported pain levels and hand 3D models and transfer them to therapists. These technologies can improve hand telerehabilitation, aid clinicians in monitoring hand conditions remotely and make decisions on appropriate therapy, medication, and hand orthoses.