Target tracking has always been a hot research topic in the field of computer vision. Tracking-Learning-Detection (TLD) is a new algorithm for online learning tracking proposed by Zdenek Kalal. In the algorithm, the computation consuming of detection module is relatively large. To solve this problem and improve the algorithm, we proposed an online learning method to adaptively update the threshold of variance classifier, which can effectively reduce the number of target boxes, improve the real-time performance and tracking accuracy. Experiments have been conducted to compare the performance of the improved TLD with the original TLD. The experimental results show that the improved TLD has better real-time performance and higher accuracy for tracking.