Markerless Performance Capture for Automated Functional Movement Screening
Abstract
Functional Movement Screening (FMS) systems are essential evaluation methods for fit-for-duty assessments and injury prevention training of warfighters and high performance athletes. With musculoskeletal injury and fatigue in general being one of the leading causes of lost duty time and morbidity, FMS screensings are widely adopted due to their accuracy, low cost, and deployability. State-of-the-art athlete analytics systems (such as the commercial Dartfish software) do not scale with the number of subjects as they rely on tedious manual annotations of body parts and joint angles for every session. Furthermore, since these manual tasks depend on the experience of the user and the individual s anthropometry, a quantifiable and consistent method for FMS scoring can be difficult. Despite over two decades of a research in vision-based estimation of 3D body poses and tracking from images and video, all existing automatic methods are struggling with the wide variations of human appearances in unconstrained environments (shape, clothing, illumination, occlusions, etc.) and only very coarse approximations of joint positions can be computed and successful body part detections cannot always be guaranteed. Our goal is to develop a comprehensive and highly deployable markerless performance capture system for Automated FMS to support US Marines training. We will use low cost video recordings (statically positioned HD webcams and mobile devices) as input and describe a 3D pose estimation framework with minimal to no user interaction. Our system will be designed to operate in fully unconstrained environments, and we plan to support fully automatic 3D landmark detection, foreground-background segmentation, and accurate biomechanical joint angle measurements even when certain body parts are occluded. We will begin our investigation with state-of-the-art convolutional network (ConvNet) with our own "US Marines" training data that we will produce in simulated environments (depth sensor and multi-view capture settings), and estimate their 3D body poses for FMS screenings. We will assess the quality of the biometric data of our method on specific FMS-based movement patterns (deep squat, push-ups, hurdle step, etc.) and compare its performance to existing depth sensor-based body tracking systems and manually annotated software solutions. We will further our explorations by generalizing our method to single view acquisitions, and handle more challenging scenarios such arbitrary body poses and shaky hand held recordings from smartphone cameras. Our research will serve as a stepstone toward fully unobtrusive tracking and interpreting gestures and actions of soldiers and civilians in training scenarios, real field settings, such as combat operations and disaster relieve, as well as security/surveillance applications around US installations to monitor suspicious activities.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 12, 2016
- Source ID
- N000141512639
Entities
People
- Hao Li
Organizations
- Office of Naval Research
- United States Navy
- University of Southern California