Abstract:
Cerebral palsy (CP) is the most common physical childhood disability, a ecting 1 in 500 people in developed western countries. The disorder is typically only diagnosed in children around 2 to 3 years of age but the disorder can also be detected in infancy by observing general movements (GM). GMs are spontaneous movements expressed by infants from the ages of 0 to 5 months old. When judged using the General Movements Assessment (GMA), risk of CP development can be predicted. The aim of this pilot study was to rstly implement a non-contact process of extracting infant kinematics using markerless motion capture techniques, and secondly, demonstrate that the capture and reconstruction of markerless motion data and inverse kinematics (IK) performed on such data can produce joint angles of equal to or less than 5 of di erence to the same motion captured using the gold-standard of passive marker optical motion capture. To address the aims, four healthy infants between the ages of 15 to 24 weeks were recruited. Infants were recorded using a passive marker optical motion capture system (Vicon) and a novel markerless method developed in this study. The 3D motion data captured by the Vicon system was a benchmark goal for the developed markerless system. The markerless method employed stereophotogrammetric techniques to nd 3D coordinates from RGB video recorded on iPad cameras. Musculoskeletal modelling software OpenSim was used to calculate the joint angles from both methods using an inverse kinematic (IK) algorithm. The mean root-mean-squared error (RMSE) and Pearson's correlation coe cients were calculated to compare the markerless method against the gold-standard. Two of the four infants returned motion capture data appropriate for comparison. Infant-4 returned the lowest mean RMSE values between the two infants at 3.07 and 4.02 for lumbar bending and extension respectively, with associated correlation coe cients of 0.75 and 0.63. 18 of the 22 joints had correlation coe cients of equal to or greater than 0.5. Analysis of the IK results show that the markerless method, developed in response to the rst aim of this study, can produce IK results within 5 or fewer of the gold-standard. The method is limited by an inaccurate biomechanical model. Further work is required to generate an accurate infant model, improve and optimise the current method, and to classify GMs. The results of this study are promising and set a foundation for the development of an automated GMA classi cation tool in the future.