Research On Algorithm Of Online Image Photometric Calibration For Visual SLAM

In recent years,with the more and more important role of autonomous driving,UAV,autonomous navigation robot and augmented reality playing in people's life,visual SLAM,as an important part of them,has attracted more researchers' attention.The visual SLAM,especially the direct method of it,largely depends on the brightness constancy assumption,that is,the gray values of a spatial point at different images are the same.However,due to the influence of scene illumination and camera photometric parameters,the brightness constancy assumption is often not fully tenable,and using such images directly will greatly reduce the accuracy of visual SLAM.Therefore,it is necessary to consider photometric calibration of image in the research of visual SLAM.In this respect,the off-line photometric calibration algorithms have been studied for a long time,but these algorithms are time-consuming and not suitable for visual SLAM,while the online photometric calibration algorithms for images are not perfect,and there is still no good solution for them,which needs further research and optimization.This thesis aims to design a more accurate and robust online image photometric calibration system based on the current photometric calibration algorithms.To this end,three aspects of work have been completed:(1)in the process of estimating the exposure time,RANSAC and multi-frame photometric information fusion(fusing the photometric information of the same spatial point in the multi frames)algorithms are used to reduce the influence of incorrect optical flow matching and photometric parameter errors on the exposure time estimation;(2)a key frame selection strategy is designed for online photometric calibration system based on the algorithm of key frame strategy in visual SLAM system and the change of photometric parameters;(3)a joint optimization algorithm of sliding window and the mature points is designed,which not only considers the pixel points in the sliding window,but also adds the previously optimized mature points as constraints to the optimization function,so as to solve the inconsistency of the estimated vignetting function in the real-time calibration process caused by only using the pixels in the sliding window during each optimization in the current algorithms.We design experiments on TUM-mono and EUROC to verify the above algorithms and test the system designed in this thesis.The results of these experiments show that theproposed methods can effectively improve the performance of the photometric calibration system.Compared with the existing photometric calibration system,the system designed can achieve better photometric calibration,so as to more effectively improve the accuracy and robustness of the direct visual SLAM.