Research On Unmanned Aerial Vehicle(UAV) Detection Technology Based On Multi-view Feature Fusion

With the rapid development of the unmanned aerial vehicle(UAV)industry,the number of civilian drones has been increasing,which leads to the phenomenon of UAV "black flying " and " indiscriminate flying " increasingly.Effective supervision of UAV has become an urgent task,and UAV target detection technology is the key to achieving effective supervision.In this paper,the research on passive radar UAV detection technology is carried out,and the main work and research results are as follows:(1)The target detection performance of common passive radar signals is analyzed and compared,and the related standards and characteristics of DTMB(Digital Terrestrial Multimedia Broadcast)signals are introduced in detail.At the same time,the passive radar signal model and the rotator UAV micro-motion model are analyzed and established,which provide theoretical sources for the future research of UAV detection in this paper.Then,two commonly used neural network structures,convolution and residuals,are introduced,which provide the theoretical basis for subsequent multi-view selection and multi-view feature fusion in this paper.(2)A UAV micro-motion multi-view model construction method is proposed.Firstly,the two principles to be followed for multi-view data construction,namely the consistency principle and the complementarity principle,are theoretically analyzed.Then,three commonly used signal time-frequency analysis methods,namely STFT(Short-Time Fourier Transform),continuous wavelet transform CWT(Continuous Wavelet Transform)and Wegener-Weil distribution WVD(Wigner-Ville Distribution),are used to characterize the UAV micro-motion and explore the time-frequency characteristics of the UAV echo signal.Then multi-view data are constructed based on the principle of multi-view construction and three time-frequency transformation methods,i.e.,the time-domain map,frequency-domain map,STFT map,CWT map,and WVD map of the signal.Finally,the single-view quality evaluation and the inter-attempt consistency complementarity evaluation are carried out by the deep learning method,i.e.,the network structures based on single-view UAV detection and two-view feature fusion UAV detection are constructed respectively by using convolutional neural networks,and the UAV detection performance of different views and the UAV detection performance of two-two view fusion are compared by experiments,and finally the view with excellent performance of immediate domain map,STFT graph,CWT graph,and WVD graph namely as the multi-view data in this paper.(3)Two multi-view fusion UAV detection methods based on data-level and feature-level are proposed.For the limitations of single-view UAV detection methods,two multi-view fusion UAV detection models based on data-level and feature-level are constructed using residual neural networks,in which the data-level multi-view fusion UAV detection model uses channel stacking to stitch and fuse multi-view data at the input side of the network,and the feature-level multi-view fusion UAV detection model first performs feature extraction by training a deep The feature-level based multi-view fusion UAV detection model first performs feature extraction on the multi-view data separately by training a neural network,and then performs feature fusion with the extracted features followed by a channel dimensional stitching operation.The experimental results show that both multi-view fusion UAV detection models can effectively improve the UAV detection accuracy,among which,the data-level multi-view fusion UAV detection model improves the average accuracy by 1.16% and the highest accuracy by 1.11% compared with the best-performing view in the single-view model;the feature-level multi-view fusion model improves the average accuracy by 1.16% and the highest accuracy by 1.11% compared with the best-performing view in the single-view model.The average accuracy is improved by 2.42% and the highest accuracy is improved by 2.11%.