This paper proposes an effective online training algorithm which resizes bounding box for visual object tracking in Convolutional Neural Network (CNN). The proposed algorithm employs partitioned bounding boxes for resizing bounding box. Compared to previous algorithm where bounding box size is rarely changed, the proposed algorithm improves the efficiency in object tracking by resizing the size of bounding boxes. The proposed algorithm is composed of three classifiers; The front-end classifier is for tracking target object by using the entire feature of previous bounding box, and the other two back-end classifiers are for resizing bounding box when the score acquired by the first classifier is lower than threshold or the number of frame exceeds a determined number. Training data for front-end classifier are extracted from previous raw result, and training data for back-end classifiers king a target in a sequence, the proposed algorithm makes more accurate result and training data of various sizes than previous algorithm by resizing bounding box. Experimental results show that the success rate and the precision for visual object tracking are improved by 3% and 5%, respectively, when compared with previous works.