LVQ-based video object segmentation through combination of spatial and color features

This paper proposes semi-automatic video object segmentation using learning vector quantization (LVQ). For each video frame, we use 5-D feature vectors whose components are spatial information in pixel coordinates and color information in YUV color space. First, the object of interest and its background are defined with human assistance. Both the object of interest and its background are then used to train LVQ codebook vectors to approximate the object shape. Next, the LVQ codebook vectors are used to segment the object of interest automatically for subsequent frames. We introduce a variable weight K for scaling 5-D vector to adjust the balance between spatial and color information for accurate segmentation. Experimental results show that the proposed algorithm is useful for tracking an object moving at moderate speed.