Incremental Learning of Weighted Tensor Subspace for Visual Tracking

Tensor analysis has been widely utilized in image-related machine learning applications, which has preferable performance over the vector-based approaches for its capability of holding the spatial structure information in some research field. The traditional tensor representation only includes the intensity values, which is sensitive to illumination variation. For this purpose, a weighted tensor subspace (WTS) is defined as object descriptor by combining the Retinex image with the original image. Then, an incremental learning algorithm is developed for WTS to adapt to the appearance change during the tracking. The proposed method could learn the lightness changing incrementally and get robust tracking performance under various luminance conditions. The experimental results illustrate the effectiveness of the proposed visual tracking Scheme.