Abstract: Heart rate detection, as a vital physiological parameter, plays an important role in medical
care, criminal investigation andinformation security, etc. Current studies on computer vision areas
have shown that heart rate signals can be obtained from videos captured by a normal webcam. The
current method can achieve relatively more desirable results in ideal experimental environments,
while the robustness of it is poorer in natural conditions when there is head shaking, noise and
shadow. In this study, we captured the region of interest by detecting the face landmarks, to reduce the
interference of the detection errors caused by the head shaking. And based on low-rank and sparse
matrix decomposition, this paper proposes a non-touch heart rate estimation model to denoise the
blood volume pulse (BVP) signal matrix in the frequency domain, so as to tackle the problem arising
from capturing heart rate signals by cameras in a non-touch way. We tested our model on the dataset
of MAHNOB-HCI and the results showed that the proposed model outperforms with 3.25% error
ratio means.

Key words: low-rank and sparse matrix decomposition, non-touch, heart rate estimation, face
land-mark detection, noise, robustness