In order to track target accurately during a long term in complicated environment, an adaptive scale feature compressed tracking algorithm is presented. A number of scanning windows with different scales and positions were obtained by construction constraint sampling. To reduce the feature dimension and improve the processing speed, the sparse random perceived matrices of different scales which can be easily computed offline were adopted to extract the features of different sampling image patches with relevant scales online. The sampling patch having a maximal classification score was regarded as the new tracking result by classifying the compressing feature via a naive bayes classifier and updating the parameters through online learning, which can realize the adaptive update of tracking location and scales. Experimental results show that the algorithm can adapt itself to the basic attitude and scale change, which is robust and does not depend on the scale selection of the initial tracking area.