Rasearch On Camera Pose Estimation Based On Unsupervised Learning

SLAM(Simultaneous Localization and Mapping)is a very important research topic in computer vision,which plays an important role in the scene of UAV,robot navigation,automatic driving and so on.Among many methods of SLAM,monocular method has become a hot research topic in recent years because of its portable acquisition equipment,low cost and not limited by the size of the scene.Camera pose estimation is a core module in monocular SLAM and plays an important role in map construction.In recent years,deep learning has developed rapidly,and it is widely used in image processing,target tracking and other fields.The application of deep learning to monocular SLAM research has gradually become a trend of development.This paper mainly uses the unsupervised learning way to learn camera pose in monocular SLAM,optimized the existing deep learning network's loss function and improved the accuracy of camera pose estimation.In addition,this paper proposes a method of generating global pose from relative pose between frames.The main work and innovation of this paper are as follows:(1)Aiming at the problem that using the loss function based on traditional inter-frame illumination error is easy to fall into the local optimal solution when estimating the camera pose,the concept of gradient illumination error is proposed in this paper.The gradient illumination error term is added to the traditional inter-frame illumination error,which improves the ability of the network to estimate the camera pose.In the process of solving camera pose by direct method,the global gradient of cost function is mainly related to two factors.The gradient of image and the derivative of coordinates of points after pose transformation.When the image gradient is small,it will lead to smaller global gradient,which will easily lead to the loss of global gradient in the iteration process,thus falling into the local optimal solution.In order to improve the disadvantage,this paper adds the illumination intensity error of points with large gradient value to the traditional inter-frame illumination error,so as to increase the guiding effect of image gradient on the overall gradient.Experiments show that by training on KITTI dataset and testing with 09 and 10 sequences,the ATE errors of camera pose are reduced by 21.4%and 33.0%respectively compared with Zhou's result using inter-frame illumination error in 201 7 years.(2)To solve the problem of camera pose's accumulating errors,an inter-frame illumination error based on allocation strategy is proposed,which provides stronger inter-frame constraints for the network.If the neural network can effectively learn the camera pose,the error of illumination intensity in projecting any two frames of the image to another frame should be minimized by the pose transformation.Based on the idea of error allocation,this paper divides the total illumination intensity errors into the sum of the previous frame,the intermediate frame and the latter frame,which provides stronger constraints for the network and further improves the ability of the network to estimate the camera pose.The experimental results show that on the 09 and 10 sequences of KITTI dataset,the ATE errors of camera pose are reduced by 4.7%and 5.7%respectively by comparing the allocated illumination intensity errors with the unallocated ones.(3)Aiming at the problem of camera trajectory visualization,a method of generating global pose from relative pose between frames is proposed.In order to display the position of the camera in the global coordinate system more intuitively,this paper adopts the idea of dynamic programming,estimates the global pose of the k+1 frame with the optimized global pose of the k-1 and k frame and the output relative pose of the k frame respectively,and then uses the quaternion spherical interpolation to optimize the global pose of the k+1 frame.Experiments show the effectiveness of the algorithm.