It is crucial to establish an effective online model for robust tracking. As existing online learning tracking algorithms do not judge whether the objective observation information is effective, a simple and efficient solution is proposed. The positive and negative samples are applied to build online object model, then feature information is extracted from the multi-scale image feature space by compressive sensing to represent object, the random fern classifier is adopted to classify and determine the online update rate by a confidence measure strategy of features. The online object model will output the sample with the highest confidence, which is decided whether to update by an shelter feedback mecanism. Experimental results show that the proposed algorithm can complete the robust tracking under the condition of long-time occlusion, illumination changing, the video sequence of 320 pixel×240 pixel, the processing speed can keep 30~50 frame/s, which meets real-time application requirement.