Research On Camera Pose Estimation On Deep Learning

In recent years,with the rapid development of artificial intelligence in China,the research of driverless and robots under the background of smart cities has become a new research hotspot.Visual SLAM is a very important research content in this field.Among them,the camera pose estimation is the core of the visual SLAM.Camera pose estimation refers to the estimation of changes in camera motion between images in different scenes or objectives.In this process,problems such as rotation between images,light intensity changes,motion amplitude changes,and sparse texture will greatly affect the accuracy of camera pose estimation,and then affect the localization and mapping of unmanned technology and robots in the environment.This paper divides the reasons that affect the camera pose estimation into two factors,the internal factor is the impact of the camera's own motion,and the external factor is the influence of external environmental factors on the feature processing.In particular,the camera pose estimation is studied in large-scale motion transformation and multiple types of indoor environments with deep learning methods,which helps the visual SLAM to stably complete the localization and mapping.The work of this article is as follows:(1)Aiming at the large-scale motion estimation problem of the camera in 3D space,this article proposes a motion transform estimation method based on dense features.This method treats the source image and the target image as a whole,and calculates the feature similarity of the image pairs globally to predict the transformation of motion between the image pairs.The overall method first uses feature pyramids to extract feature pairs of images at different scales,then fuses feature information between image pairs through correlation layers,highlights high-similarity feature fusion values in the fusion vector,and then uses motion transformation The encoder combines the fusion information vector in each feature extraction layer to predict the motion transformation,and parameterizes the transformation.Finally,the encoder of the transformation is iterated to predict the motion transformation that occurs between the two pictures.This method effectively predicts the motion transformations that occur in multiple viewpoints of the same scene in the experiment.The prediction accuracy and the extraction of key features in the image pairing are excellent.In addition,in order to visually display the motion between image pairs,this article also makes a supplementary design to the method,which qualitatively show the motion transformation between image pairs.(2)In order to accurately estimate the camera pose transformation in a variety of challenging scenes,and improve the stability of the tracking performance of the visual SLAM system,this article uses key features in the image to match the correlation and proposes a continuous feature Tracking method.First,coarse-grained image pair features are generated through a motion estimation network.Then,the transformation is used as optimization information to help distinguish between matching information and non-matching information.Finally,the matching relation matrix is obtained by using the matching discriminant classification strategy.The effectiveness of the overall method in application scenarios is verified by applying it to the traditional SLAM method.This article designs a complete visual SLAM back-end optimization module for the method to make the method a complete visual SLAM system.The experimental results also confirm that the training based on the matching relationship effectively improves the robustness of the visual SLAM system tracking in various scenes,especially in the feature sparse and weakly textured scenes.In summary,this article designs a new image feature matching algorithm for the visual SLAM field from the perspective of the internal and external aspects that affect the pose estimation of the camera.It effectively links the features between image pairs with the motion transformation.Good performance was achieved in this challenging scenario.The research provides new ideas for the field of visual SLAM,and has positive application and research value in areas such as unmanned driving.