A Motion Distortion Resistant vSLAM System for Quadruped Robots Based on Deep Learning*

For the existing visual SLAM system relying on image quality, the image motion distortion caused by the characteristics of the quadruped robot will seriously affect the stability of feature extraction and tracking in the visual front end. In this paper, we propose a visual SLAM system robust to image motion distortion to reduce the effects of image motion blur as much as possible by introducing a motion distortion preprocessing module. Specifically, this paper uses the Laplacian gradient function to calculate the sharpness of the input image and screen the image with motion distortion, and then inputs it into the lightweight deep learning network Ghost-DeblurGAN for real-time motion distortion recovery processing. After correcting for image motion distortion, Shi-Tomas feature points are extracted in the image, and tracked using pyramidal optical flow between consecutive frames. Experiments show that the proposed method can significantly improve the feature tracking loss problem. In the campus quadruped robot datasets, it reduces the average closure error by 0.079 m compared with VINS-Fusion, which can weaken the influence of motion distortion on the system, and effectively suppress the drift of pose.