Research On Deep Visual Localization Algorithm For Sequential Images

With the rising of artificial intelligence,the rapid development and application of technologies such as autonomous driving and augmented reality in the industrial field,visual localization algorithms have huge research value and application value.Many mature visual localization algorithms are mostly based on traditional stereo geometric algorithms,and are mostly based on hand-crafted features.However,these traditional visual localization algorithms have some problems such as non-robust feature points and long time-consuming feature matching.These problems greatly restrict the application of visual localization algorithms in real-life scenarios.On the other hand,with the development of deep learning technology,it has achieved breakthrough performance in image processing tasks such as image classification,object detection,and semantic segmentation.Deep learning technology has the advantages of relatively robust features and end-to-end solution tasks.In order to improve the localization accuracy of traditional visual localization algorithms,deep learning technology is an effective and feasible solution.Therefore,this paper has important research value for the research of visual localization algorithm based on deep learning.The current localization accuracy of the visual localization algorithm is poor,and the adjacent frame image localization results are quite different.To solve these problems,the paper proposes to add timing constraints to improve the localization results.The paper further uses optical flow to improve localization accuracy.For the use of optical flow information,the paper designs two network structures,early fusion and late fusion.The early fusion network structure directly integrates RGB image and optical flow information into the network to regress the camera pose.The late fusion network structure uses the same network structure to extract the features from RGB image and optical flow information.Then the features are merged to regress the camera pose.In this paper,the indoor scene dataset-7 Scenes and the outdoor scene dataset-RobotCar are used for experimental verification.The experimental results show that the proposed method effectively improves the localization accuracy.