Abstract: Remote photoplethysmography (rPPG) is to sense small changes in the skin's colour caused by blood flow in the face. Changes in lighting and movement of the face during the actual recording may also reduce the strength of the signal. Most of the current methods only model visual cues and have made little use of physiological changes associated with emotions. The multi-task framework of EP-rPPG is employed in this paper to carry out joint reconstruction of rPPG waveforms and valence-arousal estimation. A common encoder is employed to extract spatiotemporal information from the facial video, and then the two tasks are handled by separate branches. The above framework is used for both emotion prediction and, at the same time, the attention module can be employed on the intermediate features related to emotion. Then, the obtained weights are used to reweight the features in the rPPG branch and give more importance to pulse-related responses and reduce the impact of illumination changes and motion. Only the face video is required by the model for inference. According to the leave-one-subject-out evaluation method of the DEAP dataset, the mean absolute error and Pearson correlation coefficient for rPPG reconstruction of EP-rPPG were 2.93 bpm and 0.94, respectively. The classification accuracies for valence and arousal were 93.58% and 92.43%. The UBFC-rPPG dataset had good reconstruction accuracy in the evaluation as well. The above results show that emotion-related information can serve as a good auxiliary cue for the estimation of video-based physiological signals and improve the stability of the model in all kinds of datasets.