A method of online selecting local salient subregions for object tracking in the complex environment is proposed by imitating human vision characteristic of selective attention on salient regions. Subregions are randomly sampled and selected according to center-surrounding discrimination and discrimination to background, and the temporal coherence of each subregion is evaluated by the tracking confidence. Then, stable subregions with low tracking errors are selected as the support cues to estimate object state by consistence of positions. Experimental result shows the ability to handle partial occlusion and background distractions by selecting salient subregions dynamically, which leads to more robust tracking.