Hash-based methods for nearest neighbor search has been widely used in information retrieval area, such as image retrieval, data mining and text match. The methods compress original data into low-dimensional binary codes by Hash functions, and then sort and search under Hamming distance. Therefore, the methods have the advantages of efficiency and dimension insensitivity for searching large-scale data. Currently, there is little literature that discuss on learning to online Hash-based methods for real-time dynamic streaming data. Furthermore, those methods do not discuss the update frequency of Hash functions and the stability. In order to solve the problem and improve the efficiency of learning to online Hash, the confidence interval is first designed to reduce the frequency of changing the Hash table, and the objective function is proposed to keep the Hash models as stable and convergent as possible. Compared with some related online Hash algorithms on several public large-scale datasets, the proposed method is competitive under the average accuracy and training time.