Heart rate is one of the important indicators to measure human physiological conditions.Aiming at the problems that traditional signal processing methods are vulnerable to changes in light and motion,this paper proposes a non-contact heart rate measurement method based on 3D residual attention network.This method improves the spatio- temporal network,the 3D convolutional attention module(3D-CBAM)is embedded in the spatio-temporal convolution block to enhances the neural network's channel and spatial feature extraction of video,so that the model pays more attention to areas with strong physiological signal distribution.The residual structure is introduced to improve the network performance without changing the network depth.The experimental results show that the proposed method can achieve more accurate measurement.The mean absolute error(MAE)on PURE dataset is 0.52 bpm,and the root mean square error(RMSE)is 2.42 bpm;The MAE and RMSE on the UBFC-rPPG dataset were 1.17 bpm and 3.68 bpm respectively,and the recovered remote photoplethysmography (rPPG)signal was basically close to the peak position of the standard physiological signal.