Remote heart rate estimation based on convolutional neural network and regional adaptive weighting

In this paper, a heart rate prediction algorithm based on onedimensional convolutional neural network and adaptive selection of multi-region of interest signals is proposed. Firstly, different from choosing fixed region of interest (ROI) for extracting heart beat signal and facial movements, this article selects the multiple ROIs and extracts the characteristics of Chrominance signals based on the spectrum similarity to assess the significance of heart rate signals. Then signal integration is employed to effectively improve the signal-to-noise ratio of the input signal. Secondly, the existing heart rate detection methods based on convolutional neural network mostly adopt two-dimensional, three-dimensional and other large-scale networks, with large number of parameters and slow operation speed. In this paper, a light weight one-dimensional convolutional neural network is proposed. The network takes onedimensional heart rate detected signals extracted from video as input, which not only ensures the integrity of heart rate signals but also reduces the number of network parameters. Experiments on MAHNOB-HCI database show that the proposed algorithm has achieved better performance.