Denoising Real-World Low Light Surveillance Videos Based on Trilateral Filter

In this paper, a real-world denoising method for surveillance videos based on the trilateral filter is proposed. The algorithm is faced with challenges, including the absence of “ground truth” images and the complexity of the spatio-temporal distribution of the noise signal. However, experimental results have demonstrated that noise on stationary objects in such situations can be easily eliminated by averaging neighboring frames. Consequently, effective noise removal throughout the entire video can be achieved by accurately tracking and filtering moving objects along their trajectories. The model can be broadly divided into four steps. Firstly, coarse motion vectors are obtained through bilateral motion estimation. Secondly, the error vectors are judged and corrected using an amplitude-phase filter. Thirdly, these vectors are refined by performing a full searching in a small area. Finally, the noise is removed by applying a trilateral filter along the trajectory. The effectiveness of our model has been confirmed through numerous experiments, showcasing superior performance in both visualization and quantitative testing.