Abstract
Parkinson's Disease (PD) is a neurodegenerative disorder that is often accompanied by slowness of movement (bradykinesia) or gradual reduction in the frequency and amplitude of repetitive movement (hypokinesia). There is currently no cure for PD, but early detection and treatment can slow down its progression and lead to better treatment outcomes. Vision-based approaches have been proposed for the early detection of PD using gait. Gait can be captured using appearance-based or model-based approaches. Although appearance-based gait contains comprehensive features, it is easily affected by factors such as dressing. On the other hand, model-based gait is robust against changes in dressing and external contours, but it is often too sparse to contain sufficient information. Therefore, we propose a fusion of appearance-based and model-based gait features for PD prediction. First, we extracted keypoint coordinates from gait captured in videos and modeled these keypoints as a point cloud. The silhouette images are also segmented from the videos to obtain an overall appearance representation of the subject. We then perform a binary classification of gait as normal or Parkinsonian using a novel fusion of the gait point cloud and silhouette features, obtaining AUC up to 0.87 and F1-Scores up to 0.82 (precision: 0.85, recall: 0.80).