The pupil auto-tracking system is a key component of the full-automatic perimeter. Taking the advantage of integral image in counting characteristic value rapidly, we studied the relationship between training stages and total error detection rate based on the training of Adaboost strong classifier. Besides, a testing strategy of amplification detection window was introduced, and a multi-stage cascaded eye classifier for eye detection was proposed finally. It kept the same detection rate as the commonly trained strong classifier with a much lower error detection rate. In the meantime, the present article explaines the main arithmetic implement functions, as well as designs the motion control program for the jaw bracket system.