Research On Mobile Robot SLAM System Based On Monocular Vision

In recent years,with the improvement of people's standard of living,the progress of science and technology,navigation and positioning related technology gradually from the aerospace and military industry into people's lives.The original research on navigation and positioning focused on large outdoor scenes dominated by the Global Positioning System(GPS)and high-precision inertial measurement units.Nowadays,as the demand for indoor positioning navigation increases,the research of positioning navigation is moving away from GPS to indoor positioning dominated by a variety of small and inexpensive sensors(e.g.cameras,inertial sensing units,LiDAR,etc.).Simultaneous localization and mapping(SLAM)was thus proposed.Among the various sensors,the low price of cameras,the richness of information they contain,and the prosperity of the computer vision field have made people realize the value of camera-based vision SLAM research.Visual SLAM methods are divided into two main categories according to their front ends,feature point visual SLAM and direct visual SLAM,both of which have their own unique disadvantages but can complement each other to some extent,direct method is not texture sensitive,but feature point texture sensitive.The feature point method allows effective repositioning and loopback detection due to the distinguishability of the features,while the direct method does not.Visual SLAM also has the unique disadvantage that it is light sensitive,and when there is not enough light,the positioning accuracy and robustness of both direct method visual SLAM and feature point method visual SLAM will be greatly lost.Based on the above analysis,this paper mainly solves the problem of visual SLAM light sensitivity and texture sensitivity,we proposes the following solutions.First,a constant model with an error detection module is proposed based on the SVO algorithm framework,which allows SVO to have an initial value closer to the true value before estimating the posture.A better initial value not only improves the accuracy of the algorithm and avoids non-convexity to some extent,but also improves time efficiency by reducing the number of iterations because the initial value of its iteration is closer to the true value.We then propose a very effective preprocess method based on VI-ORB for light problems.We first perform a gamma change on the image to increase the contrast in the dark areas of the image,but since the gamma change increases the contrast in the dark areas while decreasing the contrast in the bright areas,we take the contrast-limiting grayscale histogram equalization method to process the gamma-transformed image to increase the contrast of the image as much as possible without increasing noise.In order to solve the problem of sparse textured VI-ORB scenes,we introduced the light flow tracking method and improved the system frame management and map point management,so that VI-ORB can still run robustly in low-light,low-textured environments.Finally,to verify the accuracy of our algorithm,we perform comparison experiments with State of the art's algorithm in the visual SLAM domain on a public dataset.To verify the robustness of the algorithm,we built a robotic experimental platform to compare the algorithm's operation in a challenging environment,and the experimental results prove that our algorithm outperforms other State of the art algorithms in most cases.