| In this paper,the single-soldier loitering munitions is used as the application platform.In order to solve the problems of small target imaging size and complex environment interference during the detection process,the research on the eagle-eye optical system is carried out based on the bionic eagle-eye vision,and An improved method is proposed based on the YOLOv4-tiny target detection algorithm.At the same time,due to the limited space of conventional ammunition,this paper builds a hardware system based on the Jetson Xavier NX embedded platform and conducts experimental tests.Firstly,the visual principle,physiological structure and visual information processing channel of the eagle eye are introduced and analyzed.Based on the research of the eagle eye vision principle,the functional structure of the bionic eagle eye system is designed using the loitering munitions as the application platform.And this paper gives the design process of the system,which establishes the foundation for the follow-up work.Secondly,in order to simulate the eagle eye’s central fovea and lateral fovea structure,a set of bionic eagle eye vision dual-path optical system was designed and aberrations were corrected.At the same time,the structure of the central foveal optical system and the lateral foveal optical system was designed and modeled,and the overall appearance structure of the bionic eagle eye system was designed based on the loitering munitions platform.Then,using the YOLOv4-tiny network model as the main algorithm of the bionic eagleeye target detection,the network structure was improved and the visual attention mechanism was introduced.The data set of this article was made and the data was enhanced by using three combat vehicle models.,Use the K-means++ algorithm to cluster and optimize the anchor frame to improve the accuracy of small target detection in complex environments,and then train the improved network model on a small server to obtain the final training weights.Two pictures which are randomly selected from the verification set are chose to test the improved algorithm.Compared with the YOLOv4-tiny algorithm,the accuracy of the improved algorithm is increased by about 1.5%,and the recall rate is increased by about 21%.Finally,a hardware system based on the Jetson Xavier NX embedded platform is built.Based on the characteristics of Gstreamer,the system hardware is used for video encoding and decoding,and CPU resources are released.Based on the characteristics of Tensor RT,the improved network model is optimized and accelerated,and then the improved network model is transplanted to the hardware system,and static experiments of target detection under different brightness levels in complex background environment and simple background environment are carried out.Then,dynamic detection experiments are carried out to verify the targets in different background environments.The detection speed is about 45 FPS,which verifies the stability of the system can be used for real-time detection of targets.In summary,this article designs and simulates the optical path and structure of the bionic eagle eye optical system,optimizes the accuracy of the bionic eagle eye target detection algorithm,and reduces the missed detection of small targets in complex environments.Improved the accuracy of target detection,and finally realized real-time target detection on the Jetson Xavier NX embedded platform. |
PDF Full Text Request