A Space Target Perception Method For Artificial Intelligence Chip

With the development of space technology both for military and business,space situation awareness is becoming more and more important and space target recognition is the key technology for space situation awareness.The main task of space target recognition is to obtain information of space targets.High resolution optical telescopes can directly obtain images of space targets and has good detection capabilities for important space targets.Nowadays,high resolution telescopes are installed both on the ground and in space platform.Because there would be very limited energy and space for space platform,it would be necessary to develop space target perception method for space targets.Based on this requirement,this article has carried out the following work: design of space target datasets,development of space target recognition technology,space target information perception technology,and the deployment of space target detection and perception networks in artificial intelligence chips:(1)In terms of space target datasets construction,due to the difficulty in obtaining real observation images with ground truth labels,there is a lack of a large amount of data to train neural network.Therefore,a pipeline for generation of the space target images is designed.The pipeline can generate high resolution images of space target with different types,distance and attitude,and these simulated images can have different blur and noise levels.(2)In terms of space target recognition,the space target detection network based on YOLOv3 is studied.To solve the problem that the space target detection network can only identify space targets with a certain degree of degradation,a space target load detection network based on hybrid training is designed.Experiments show that the recognition accuracy of our space target detection network can detect multiple space targets with different loads under different degrees of blur and noise levels with more than 80% accuracy.(3)In terms of space target information perception,two space target detection networks are trained separately for different characteristics of the space target and a space target information perception expert system based on these two detection networks is designed.The output of the above two networks is merged to recognize states of the space targets.(4)In terms of the deployment of space object detection and perception networks on artificial intelligence chips,due to the limited space and energy,it is not realistic to use large and energy-consuming GPUs for space object perception.A real-time spatial object perceptron with strong performance and suitable for edge computing is necessary.This paper proposes to deploy the space target detection and perception network model on the NVIDIA Jetson AGX Xavier device to realize the real-time detection and perception of space targets.