HRV-Transformer: Transformer-based non-contact heart rate detection

As a kind of physiological measurement utilizing a camera to sample face feature, there are two important factors: accuracy and real time. This paper presents a framework based on Transformer called HRV-Transformer. The framework takes full advantage of Transformer, which improves the accuracy of non-contact heart rate detection and accelerates the speed of model training. However, the Transformer structure has a pronounced disadvantage which is redundancy of the structure. Through the removal of some non-essential modules, the performance of Transformer is not affected, thus realizing the real-time heart rate detection. In order to further improve the accuracy and latency, the original video frame is compressed into frames and frames difference, which can decrease the number of parameters during training. There are four kinds of data respectively used as the input of HRV-Transformer framework, so as to compare the effect of the different size and type of the data. Through experimental verification, compared with the traditional methods, the proposed model obviously improves the accuracy and latency of heart rate detection based on a camera.