Skip to Main Content
The paper discusses a novel unsupervised learning approach for tracking deformable objects manipulated by a robotic hand in a series of images collected by a video camera. The object of interest is automatically segmented from the initial frame in the sequence. The segmentation is treated as clustering based on color information and spatial features and an unsupervised network is employed to cluster each pixel of the initial frame. Each pixel from the clustering results is then classified as either object of interest or background and the contour of the object is identified based on this classification. Using static (color) and dynamic (motion between frames) information, the contour is then tracked with an algorithm based on neural gas networks in the sequence of images. Experiments performed under different conditions reveal that the method tracks accurately the test objects even for severe contour deformations, is fast and insensitive to smooth changes in lighting, contrast and background.