Using the traditional cuff based instruments to estimate the blood pressure (BP) values is inconvenient and difficult to have a continuous measurement. On the other hand, using the intelligent watch based instruments to perform the continuous BP estimation for monitoring the human health conditions is convenient and easy. This paper proposes a method for estimating the BP values continuously only using the photoplethysmogram (PPG). First, the PPGs are denoised by the discrete cosine transform (DCT). It is worth noting that the conventional DCT denoising approach only takes the low frequency DCT coefficients. On the other hand, this paper proposes a training based method for selecting the DCT AC coefficients. Here, the DCT AC coefficients refer to those non-DC DCT coefficients. Then, the features based on some specific points in the PPGs are extracted and the feature vectors are categriozed into two classes of the BP values using the random forest (RF) classifier. Third, for each class and each type of the BP values, three popular regressions including the support vector regression (SVR), the RF regression and the L1 norm criterion based linear regression are used to estimate the BP values. These three estimated BP values are fused together via an L2 norm based regression. It is worth noting that different classes and different types of the BP values are considered separately. Hence, each class and each type of the BP values can be estimated more accurately. Moreover, since the multi-model fusion is employed for combining these three regressions, the overall estimation results are more accurate. The computer numerical simulation results show that the average root mean squares errors (RMSEs) of the SBP and the DBP estimated by our proposed method are 9.12 and 8.19, respectively. In fact, our proposed multi-model fusion based BP estimation approach achieves a higher accuracy compared to the individual regressions.