Surveillance video denoising based on background modeling

Because of the characteristics of photoelectric sensors and the working environment of cameras, real-time surveillance video contains much noise, which does not only decrease the subjective visual quality, but also increases the output bitrate of video encoder. The effect of partial spatio-temporal smoothing is not evident. According to the characteristics of surveillance video, we propose a novel algorithm based on video content, setting up adaptive background models to accomplish foreground segmentation, reducing background noise via model parameters and foreground noise via 3D median filter. To the sequences of "hall_monitor" polluted with Gaussian or Poisson noise, the results show that the new algorithm increases PSNR about 8 dB, and saves over 90% of encoder output bitrate.