Research On Visual Relative Position Estimation Algorithm Of Asteroid Flying Phase Based On Recurrent-Convolution Neural Network

The detection and resource utilization of asteroids in deep space is an important development direction in the field of space science.The high-precision localization and navigation of the detectors during inspection,flight and landing phase in the mission of exploration on unknown asteroids is a key problem we are currently facing.Because the period of communication is too long in deep space,the asteroid detector need be provided with a highly autonomous navigation method to ensure the implementation of tasks in the detection.The navigation method that uses the visual sensors to sense the surrounding environment information,which is called visual navigation,can obtain more feature information by taking an image with an optical camera in the process of approaching the planet,thereby improving precision of navigation.In the field of visual navigation,visual SLAM(simultaneous localization and mapping)algorithm is a hot research topic.The algorithm can not only locate relatively and mapping the scene with the correlation between visual images,but also guarantee real-time.Whether for relative localization,navigation and obstacle avoidance in exploration missions on unknown asteroid,or correction orbit for combining orbital dynamics,even for decision-making of detector,it is of great significance.In the front end of the visual SLAM algorithm,we usually use the feature point method to associate images information,which is called extraction and matching of image features,thereby realizing obtain relative pose information of camera through combining geometric relations between images.Due to the unknown terrain and the complex lighting environment on asteroids' surface,changes in illumination intensity and local shadowing may occur.These changes can affect the accuracy of extracting image features,which in turn affects the accuracy of visual localization and navigation.Therefore,we need an image feature extraction and matching algorithm that can be stable to illumination changes,thereby obtaining more accurate camera relative poses.Because the traditional algorithms for feature extraction and matching can't balance high precision and high efficiency,in the SLAM algorithm with high real-time requirements,we usually consider using some image feature algorithms that reduce accuracy and robustness to improve the calculation speed,but its stability to illumination changes may be poor.In recent years,deep learning algorithms are a research hotspot in the field of image processing.In these algorithms convolutional neural network(CNN),instead of manually extracting features,is used for improving the performance of algorithms in image processing tasks,and deep learning algorithms can be driven to learn adaptability to disturbances such as changes by a large amount of data.During the flight phase of the asteroid detector,the camera's shooting scene is similar to the aerial scene,and the homography matrix can be used for aerial scenes.Therefore,we proposed a pose estimation method based on homography matrix,in which the homography matrix estimation algorithm was based on deep learning.After learning,it was able to extract feature and estimate homography matrix for visual images of various asteroid environments.At the same time,considering the flight coherence of the aircraft subject to orbital dynamics constraints and the characteristics that recurrent neural network(RNN)can process timing signals,we used a framework based on a recurrent-convolutional neural network(RNN-CNN)for estimation of the homography matrix,which improved the accuracy of estimating the homography matrix,and improved the robustness to disturbances such as illumination changes and geometric deformations.Finally,a series of asteroid surface images that was obtained during flight phase of the detector,are generated by simulating asteroid environment and imaging.The algorithm was applied to the pose estimation of the image sequences to test the performance of the proposed algorithm.The main work of this paper is summarized as follows:1.We used the public dataset to train the model parameters of the supervised and unsupervised CNN-based homography estimation network,and test them in term of the estimation error and stability to the illumination changes.In addition,in order to obtain a model with better stability to illumination changes,we introduced illumination noise in method of data augmentation,improved the network structure,and adjusted the loss function;2.Considering the sequence of the input image,the RNN structure was built on the basis of the CNN-based homography estimation network framework to obtain an unsupervised RNN-CNN-based homography estimation network,which improved estimation accuracy and stability to changes in il umination;3.Through simulating the illumination and the shadow of asteroids' surface and the motion of the asteroid detector during the flight phase,combined with imaging simulation,a sequence of asteroid surface images were generated.The algorithm this paper proposed was used to estimate the relative pose of asteroid s' surface image sequences.The comparison of the pose estimation results from the algorithm we proposed and the relational algorithm shows that the algorithm we proposed had comprehensive advantages in real-time and accuracy.