Research On Monocular Visual Localization Of Mobile Robot Based On SLAM

With the development of artificial intelligence,the integration of mobile robots and technologies has brought great changes to the traditional industry,such as using robots with autonomous path planning to sort and transfer and self-driving vehicles with autonomous navigation to draw high-precision maps.Simultaneous localization and mapping(SLAM)technology is a key for autonomous navigation and positioning of mobile robots.Compared with traditional sensors such as laser,camera sensors equipped with real-time visual positioning and map construction have the advantages of low cost,small size and abundant information,which are widely favored by researchers.This paper improved the algorithm of front-end visual odometer.Results show that the accuracy of monocular visual positioning,real-time performance and robustness of the system of mobile robot based on SLAM achieves favorable performance.The main contents and innovations are summarized as follows:(1)The traditional methods are easy to cause a large number of features to be aggregated,also leads to mismatch or match failure during inter-frame matching.A feature selection method based on quadtree was proposed.The statistical analysis of the points tracking duration shows that the feature homogenization helps extend the tracking time.The longer the features are tracked and optimized,the higher the accuracy,and the more accurate the composition and positioning.(2)The oriented fast and rotated brief(ORB)feature does not have scale invariance,a scale was added to oriented fast and rotated brief(SORB)feature which was proposed to improve the accuracy of the system when the camera changes in distance.In addition,the previous feature matching is to perform local matching independently.It is easy to matching when there are large numbers of similar textures in the image or the camera rotating.To solve these problems,an orientation algorithm based on global consistency and local similarity was proposed.The matching algorithm added global and orientation information on the basis of local information,narrowing the search area of feature,which improved the accuracy of matching rate.We validated the effectiveness of the proposed algorithm with TUM data.(3)According to the characteristics of camera motion image,a calculation strategy based on image structural difference was proposed.Furthermore,the system accelerated running speed by reducing the numbers of matching points.The feature matching runtime was counted.The results show that the structural strategy speeds up the matching calculation.(4)We proposed a non-linear optimization method to the front-end visual odometer,which optimized the overall data.Due to the cumulative error caused by the estimative errors in the long-term mapping and estimative posture processing.At the same time,the estimative posture was initialized.It is susceptible to noise interference and improved the posture calculation step to optimize the initial estimated posture.We validated the effectiveness of the proposed algorithm with KITTI data.