Shape Prior Active Contours for Computerized Vision Based Train Rolling Stock Parts Segmentation

Abstract

Computer automation of rolling stock involves determination of individual parts to be examinated for defect Identification from the videos of a moving train. Video frame segmentation using Chan Vese active contour model (CV-AC) results in a full bogie binary image that makes impossible to track individual parts. To segment individual parts and track their shapes along the length of the train is a challenging task. It could be achieved by using shape prior seeds (SP-CV-AC) as destination contour from individual parts of the bogie for the Chan vese active contour model. Spatial distances are used to propel the initial contour towards final shape contour. The results demonstrate the quality of video segmentation algorithm based on destination seed shape priors. The quality of the proposed segmentation algorithm is computed using factual segmentation score (FSS) between shape prior and hand segmented portions of the rolling stock. Further the paper compares shape prior segmentation model with no-shape prior active contours to specify the importance of shape prior models for complex image processing tasks related to intelligent maintenance systems with computer vision.