Research On Key Technologies Of Infrared Imaging SLAM

Simultaneous Localization And Mapping(SLAM)is an important problem in computer vision.It is widely used in the fields of automatic driving,navigation,imaging guidance,and surveillance.Infrared cameras are applicable at night,in haze,and in poor illumination.Hence,infrared SLAM can be used under more circumstances than visiblelight SLAM.However,due to the difficulty in obtaining infrared images,weak image texture and low signal-to-noise ratio(SNR),there is less research on infrared SLAM in the literature.This thesis focuses on a high-precision and robust infrared SLAM system,mainly addressing the problems of infrared camera calibration,SLAM algorithm design and SLAM performance evaluation.The main contributions of this thesis are as follows:1.In order to select suitable image features for infrared SLAM,this thesis studies the performance of various features through theoretical analysis and experiments.The results show that in the infrared image,the performance of feature point is significantly deteriorated compared to that of the visible-light image,while the performance of edge feature remains almost the same.Therefore,in infrared camera self-calibration and SLAM,the edge is more suitable to be used as the image feature.In the research,the evaluation criteria of spatial distribution entropy and local matching recall-precision curve are proposed,which is more suitable to guide the feature selection in infrared SLAM.2.To get the intrinsics of infrared camera for the operation of SLAM algorithm,this thesis proposes an infrared camera self-calibration method based on edge matching.The experiments show that compared with the feature point-based self-calibration,the projection error of our method is reduced from 1.57 pixels to 0.61 pixels.Meanwhile its performance is as accurate as that of the traditional target-based calibration method,which is available to support the SLAM running.3.To improve the robustness and accuracy,an infrared SLAM algorithm based on the integration of edge and feature point is studied.The experiments show that compared with the traditional feature point-based SLAM,the successful running probability of our algorithm is improved from 37.6% to 80.4%,and the localization accuracy is also improved.In the meantime,the direct SLAM is almost completely out of service in the test.Therefore,the proposed algorithm is more suitable for infrared cameras than the traditional methods.In addition,the proposed algorithm runs at 19 FPS in 640×480 infrared videos using laptop CPU,which is able to meet the real-time requirements.4.To evaluate the localization accuracy of infrared SLAM,this thesis proposes a camera motion measurement algorithm which integrates the satellite navigation(GNSS)and visual SLAM using an optimization scheme.In the experiment,the localization error of GNSS is 3m,the average localization error of visual SLAM is 20 m,and the average localization error of the proposed algorithm is 0.82 m.The accuracy of the proposed method is significantly higher than that of GNSS or visual SLAM.Therefore,the trajectory measured by the proposed method can be used to evaluate the localization accuracy of the infrared SLAM.This research features low cost and easy operation,and does not rely on other special equipments except infrared cameras.Therefore,this thesis can be adapted to reduce the research requirement of infrared SLAM,and propose a new direction for the development of infrared SLAM.