Understanding the Influence of Random Impulse Noise on Visual SLAM

The localization accuracy of visual SLAM depends on the image quality. However, in postdisaster rescue missions, the images obtained by the camera often contain considerable noise, which affects the pose estimation based on visual SLAM. In this paper, we study the influence of random impulse noise in images on the localization accuracy of visual SLAM, and reduce these influences by denoising and removing mismatches. First, the camera image is preprocessed by the traditional image noise reduction method. Aiming at the problem of a large number of mismatches in optical flow tracking due to the influence of residual noise, the improved random sample consensus method is adopted to remove it. Preliminarily judge the correct matching probability of optical flow tracking results by normalized cross-correlation matching before random sampling. Then use guided sampling to select matching points to estimate the camera motion model, to increase the robustness of the SLAM system. Finally, our method is verified in the open-source solution VINS-Fusion. Experiments show that after random impulse noise is added to the KITTI dataset, the pose estimation accuracy of the improved SLAM is higher than the pose estimation accuracy after noise reduction only, and it is also higher than the pose estimation results of the original images in multiple sequences of the KITTI dataset.