Heart rate is an important physiological indicator of the human body. It can evaluate the activity of the heart and the degree of fatigue. It can be further used to assess emotions by monitoring heart-rate changes and voice recording or breathing. Single-lead or multilead equipment is often used to monitor electrocardiogram (ECG) in medical treatment. However, it brings inconvenience to daily life to record the heart rate from a free-living natural environment. Due to the subject's voluntary or involuntary movement, it is difficult to collect ideal photoplethysmography (PPG) signal by using wearable device (WD) as it is continually interfered by the motion artifacts (MAs). These MAs seriously affect the calculation and online monitoring of the heart rate. This article proposes a WD with real-time online fusion learning for heart PPG tracking. In particular, it proposes a recursive least-squares algorithm to fuse three-axis acceleration signals to remove MAs from PPG signals. In order to further determine the heart-rate frequency range and select the appropriate spectrum peak, a decision tree spectrum interval estimation method is proposed to provide prior information for PPG signal spectrum tracking, so as to optimize the spectrum peak estimation. The performance of the proposed device, which is an integrated Internet of Things with wearable platform, has been shown to be robust to natural and unpredictable situations. The experimental results have confirmed that the proposed method can adaptively track the PPG signal with high accuracy.