The Method Of Non-cooperative Spacecraft Pose Measurement Based On Neural Network

In recent years,the number of space missions aiming at non cooperative spacecraft is on the rise.The measurement of relative position and attitude between the target spacecraft and non cooperative spacecraft is the premise of a series of subsequent space missions.In this paper,a relative pose measurement method based on neural network is studied by using monocular camera with simple structure,mature technology and low power consumption as sensor.In this paper,the target spacecraft is approximately regarded as a rigid body,and a group of key sparse points with strong visual characteristics on the target are selected to represent the target spacecraft.Thus,the problem of relative pose measurement is transformed into the problem of key point detection and multi-point perspective.In the key point detection task,11 key points(including 8 corner points and 3 antenna endpoints)on the target spacecraft are selected as the detection objects.Finally,the random sampling consensus(RANSAC)and Levenberg Marquardt(LM)algorithm is used to solve the PnP problem and obtain the relative pose.The key point detection process of the target spacecraft is divided into two stages: the first stage is to capture the image region containing the target spacecraft.Considering that the scale of the target spacecraft varies widely in the field of view,most of the pixels are useless redundant information.In order not to increase the computational complexity of the key point detection network model,a target detection network is used to segment the target area.In this stage,the accuracy of the target detection network model is not high,and the emphasis is on real-time.In this paper,the representative yolo algorithm based on single-stage target detection network is used to design a target detection network suitable for this task.The inference time of a single image is only 8 ms,and the real-time performance is good.The second stage is to detect key points from the intercepted target area.In order to extract the key points from the local image of the detected target with higher accuracy,this paper designs the key point detection network with RESNET network as the backbone network and multi-layer deconvolution module as the high-resolution feature map generation network.Through the test on the data set,the prediction accuracy,inference time and model size of the models under different backbone networks are compared.The results show that the network with resnet101 as the backbone has better comprehensive benefits.The training of the target detection network and the key point detection network needs the target bounding box and the pixel coordinates of the key points on the image plane to generate labels.In order to avoid the uncertainty of artificial accidental error introduced by manual labeling,which will affect the prediction accuracy of the trained network,this paper uses the idea of multi view triangulation to transform it into an optimization problem,The goal is to minimize the re projection error.Particle swarm optimization algorithm is used to solve the problem.After obtaining the three-dimensional coordinates of these key points,the key points are re projected to the image plane combined with the true value pose tags of each image,and the pixel coordinates of the key points on each image are obtained.The heat map is generated by two-dimensional Gaussian distribution as the key point detection tag,The smallest rectangle surrounding these 11 points is slightly enlarged and used as the label of the target detection network.When some of the key points are invisible or occluded,the detection results of these key points can not be used for pose calculation.When generating the tag heat map corresponding to the invisible key points,the pixel values are all set to 0.When searching the predicted heat map,the invisible points can be identified and eliminated by setting a threshold.For the key points occluded by the target spacecraft itself,In the same way,by setting a reasonable threshold,we can first remove a part of the prediction heat map,and then embed the iterative algorithm into the random sampling consistent algorithm,which can further remove the unavailable key points and ensure the accuracy of the results as much as possible.