An automated form of video image analysis applied to classification of movement disorders

Video image analysis is able to provide quantitative data on postural and m ovement abnormalities and thus has an important application in neurological diagnosis and management. The conventional techniques require patients to be videotaped while wearing markers in a highly structured laboratory envir onment. This restricts the utility of video in routine clinical practise. W e have begun development of intelligent software which aims to provide a mo re flexible system able to quantify human posture and movement directly fro m whole-body images without markers and in an unstructured environment. The steps involved are to extract complete human profiles from video frames, t o fit skeletal frameworks to the profiles and derive joint angles and swing . distances. By this means a given posture is reduced to a set df basic par ameters that can provide input to a neural network classifier. To test the system's performance we videotaped patients with dopa-responsive Parkinsoni sm and age-matched normals during several gait cycles, to yield 61 patient and 49 normal postures. These postures were reduced to their basic paramete rs and fed to the neural network classifier in various combinations. The op timal parameter sets (consisting of both swing distances and joint angles) yielded successful classification of normals and patients with an accuracy above 90 %. This result demonstrated the feasibility of the approach. The t echnique has the potential to guide clinicians on the relative sensitivity of specific postural/gait features in diagnosis. Future studies will aim to improve the robustness of the system in providing accurate parameter estim ates from subjects wearing a range of clothing, and to further improve disc rimination by incorporating more stages of the gait cycle into the analysis .