MOVIT: MONOCULAR VISION-BASED TRACKING

Abstract

Cerebral Palsy (CP) is a physical disability that affects approximately 17 million individuals globally. CP can severely impact the development of motor, cognitive, and social skills. Recent research efforts in this domain have led to the development of a series of assistive robot systems designed for crawling-age infants (aged 4-11 months) who are at risk for CP and related motor disorders. These robot systems provide early intervention to mitigate the effects of the above motor disorders. The robot systems capture and interpret infant limb motion in 3D and physically move an infant in response to meaningful crawling-like limb motion. Inertial measurement units (IMUs) are used for the motion capture (mocap) process. IMUs are highly sensitive to electromagnetic fields. Consequently, the presence of electromagnetic interference (EMI) sources in the surroundings causes the assistive robots to malfunction.

Thus the research problem is posed as follows. There is a need for the development of a new mocap approach to replace or augment the existing mocap system for infants. The key requirements are that crawling motions of infants should be captured and the approach must not be sensitive to EMI. The research scope is limited to tracking motion in 3D and does not include methods for automatic gesture recognition or classification. There are two research questions: 1)~What are the requirements for capturing crawling motions of infants? 2)~To what extent does a mocap system not subject to EMI, meet the above requirements?

The contributions of this research are as follows. Quantitative data on infant crawling motion from past works have been collected and presented in a form useful for the design of mocap systems. A novel approach for mocap based on planar pattern vision markers has been developed. The effects of changing various design parameters on the tracking accuracy has been documented on the basis of physical tests. A performance model has been developed to predict tracking accuracy based on the various design parameters and to allow for comparison with other systems based on tracking planar pattern vision markers.

Key conclusions of this research are as follows. The magnitude of the smallest meaningful crawling motion that an infant can make is 74.6~mm. The worst-case tracking error for the developed system is 19.9~mm. Further evaluation needs to be done to determine whether this is practical for existing gesture recognition and filtering methods.