| The construction of airborne wireless optical communication system on light and small UAV platforms can realize high-speed and stable data transmission,so it has become a key research focus and hot spot for domestic and foreign personnel.The most challenging technical difficulty in wireless optical communication link construction is how to achieve fast and accurate capture,alignment and tracking.In order to achieve fast capture and tracking and localization of target signal spot,this paper is based on GPU-accelerated neuron network algorithm for the construction of wireless optical communication link,and the following research is conducted:First,this paper introduces the initial pointing,capturing mode and scanning method commonly used in the process of wireless optical communication link establishment.Based on the beacon speckle localization algorithm,the computational principles of the traditional speckle localization algorithm and the image recognition-based convolutional neural network algorithm are described.By comparing the results of the two computational methods,it is concluded that the convolutional neural network is more accurate and less susceptible to interference in the complex background for beacon spot recognition.Therefore,in order to improve the success rate of communication link establishment,this paper applies the target detection algorithm in the field of deep learning to the capture and tracking of beacon spots in wireless optical communication system for further research.In order to make the convolutional neural network run on embedded devices with limited load and power consumption,and to improve the operation speed and accuracy,the existing structure must be optimized and improved.For the neural network structure improvement,the main network structure uses a lightweight neural network Shuffle Net,Mobile One for the design combination,which enhances the feature reuse of the shallow feature information.The new spatial pooling pyramid structure SPPFCSPC structure is used in the pooling layer part,which can solve the extraction problem of image feature reuse and the distortion problem caused by image scaling and cropping and increase the target recognition accuracy.In the detection head part,we use the Decouple Head,which is the latest public research result,to improve the fitting speed of the neural network during training and the target detection speed during inference.In order to obtain higher detection accuracy,nine representative anchor frames are selected by comparing and analyzing the Kmeans and Kmeans++algorithms in the anchor frame re-clustering experiment with the fiducial spot data set.For model performance evaluation,the authoritative MSCOCO dataset was used as the evaluation standard for accuracy evaluation,and Jetson Xavier NX was used as the runtime platform to examine the runtime speed for speed evaluation.The results show that the improved neural network improves by 18.2%on m AP@0.5 and 27.4%on m AP@0.5:0.95.In addition,in order to be able to further improve the running speed,this paper performs GPU engineering acceleration on the trained network model,and the results show that it achieves 117.6 FPS in terms of running speed,which is 112 FPS higher than that in the CPU running case.Finally,a first-class active optical tracking system is designed based on a GPU-accelerated improved neuronal network algorithm,and the system is analyzed by PID mathematical modeling applicable to light and small UAV airborne platforms.Different indoor experiments are built in the laboratory to analyze and verify the tracking performance and tracking error of the wireless light capture tracking system,in which the static tracking error is 37.92urad,the dynamic tracking error is 46.3urad,the maximum tracking angular velocity is 5°/s,the maximum tracking angular acceleration is 224°/s~2,the azimuthal dynamic angular accuracy wasσ_A=0°0’30.24’’,the pitch The dynamic angle measurement accuracy isσ_E=0°0’40.68’’. |
PDF Full Text Request