In dynamic head scenes, current non-contact blood oxygen saturation measurement methods have low accuracy. To solve this problem, we propose a denoising method based on improved adaptive noise complete set empirical mode decomposition and wavelet threshold. This method aims to extract pulse wave signals with a high signal-to-noise ratio. Firstly, in order to solve the problem of false components and mode aliasing in the early stage of decomposition and reconstruction, white Gaussian noise is added to the decomposition process to make it become an improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN), to reduce the residual noise in the modal components. Then, ICEEMDAN is used to perform mode decomposition of pulse wave signals of red and blue channels. The db8 wavelet basis function is used to perform 3-stage decomposition and reconstruction on components within the blood oxygen spectrum range. The reconstructed signals are used for subsequent calculation of blood oxygen value. Finally, the experimental comparison and analysis of the blood oxygen saturation results measured in different dynamic head scenes show that the average error of blood oxygen saturation obtained in different head scenes is 0.73%, which is 1.93% lower than the average error of other algorithms. The denoising method proposed in this paper has good stability in different head scenes and can meet the needs of daily blood oxygen saturation measurement.