A novel online clustering-based object localization strategy using learnings and human interest priors

Moving object localization is a popular study in contemporary computer vision, provided the fact that many challenging problems, such as illumination changes, obstacles, object shape transformation, etc, are still to be deeply investigated and properly tackled in moving object localization for the time being. In this study, a new clustering-based strategy is introduced to realize the moving object localization task for the first time. The strategy is divided into two steps. First, a spatially-weighted isotropically-scaled Gaussian measure is adopted to reveal the pixel-wise similarity in each individual frame of a video clip or image sequence. This similarity measure aims to differentiate pixels belonging to the group of moving target object and others. A supervised spectral clustering technique is incorporated to determine the parametric form of the similarity measure, which is associated with the idea of incremental learning. Second, an unsupervised out-of-sample extension is incorporated to finalize the differentiation of pixels, in order to obtain localization outcomes, for the consideration of efficiency boosting. Our strategy has been evaluated based on a database composed of nearly 1000 frames, and compared with several well-known approaches in this moving object localization task. Promising results are demonstrated based on a series of statistical tests. It is suggested that the newly introduced strategy is superior towards other compared methods in this study.