In order to improve the tracking accuracy of active radar/infrared detector for boost-glide aircraft and give full play to the advantages of various types of observation information，a heterogeneous information fusion tracking algorithm based on multi-scale model is proposed.Aiming at the problem of different sampling frequencies of infrared detectors and active radars，a multi-scale model is constructed for the detection system.In the fast-scale model，according to the observation data of the infrared detector，the Unscented Kalman Filter(UKF) is used to complete preliminary estimation of the target state.In the slow-scale model，accurate estimation of the target state is completed by using the fused active radar observation data and the Converted Measurement Kalman Filter (CMKF).Simulation results show that the proposed algorithm can achieve more accurate tracking effect than the traditional algorithm.

In order to solve the UAV path planning problem in complex three-dimensional environment，the traditional Dung Bettle Optimizer (DBO) algorithm and Grey Wolf Optimizer (GWO) algorithm are combined to propose a DBO-GWO algorithm.Firstly，the algorithm combines chaos strategy with quasi-reflection strategy to generate the initial population to fully explore the solution space.Secondly，the foraging behavior of the DBO algorithm and the position update strategy of the GWO algorithm are combined to help the population get rid of the local optimum and further improve the performance of the algorithm.Finally，Levy flight strategy and dynamic weight coefficient improvement algorithm are adopted to improve the diversity of the population in the later stage of the algorithm，and balance the exploration ability and the convergence speed of the algorithm. By testing with benchmark functions and conducting simulation experiments，the results demonstrate that the proposed algorithm shows faster convergence speed and the ability to generate better paths in the same environment，which shows that the proposed algorithm can effectively solve the path planning problem of UAVs.

Aiming at the problems of time-varying path constraints and unknown external interferences in unmanned helicopter formation system，this paper proposes a cooperative safety tracking control method for unmanned helicopters based on boundary protection algorithm.Firstly，for time-varying path constraints caused by external geographical environment，a new safety expectation path is designed based on a class of boundary protection algorithm.In order to counteract the impact of unknown external interferences，a disturbance observer is adopted for estimation.Then，a safety tracking controller is designed by combining backstepping control method，and the convergence of all signals in closed-loop system is proved by Lyapunov stability theory，which ensures the safety tracking performance of the unmanned helicopter formation system.Finally，the effectiveness of the proposed control method is verified by numerical simulation.

Aiming at the problem that it is difficult to reflect the battlefield situation game situation in dynamic Weapon-Target Assignment(WTA)，a Tent Random Particle Swarm Optimization WTA algorithm based on Deep Q Network (DQN) is proposed，namely TRPSO-DQN.Firstly，a DQN is used to generate a game matrix based on the battlefield situation，and then，an improved particle swarm algorithm is used to solve the Nash equilibrium，so as to solve the problems that the traditional linear programming results are inaccurate and the large dimensional matrix cannot be solved.Finally，the proposed TRPSO-DQN algorithm is used to realize Dynamic WTA(DWTA).Experiments show that the algorithm is highly adversarial in realizing DWTA，the results of WTA are more reasonable，the convergence speed of the algorithm is faster，and the Nash equilibrium solution accuracy is better than other algorithms.

A UAV target tracking algorithm that combining Siamese network with Transformer are proposed to address the issues of high computational complexity and low tracking accuracy in traditional target tracking algorithms，effectively addressing various challenges during the UAV tracking.Firstly，a backbone network is designed based on the lightweight feature extraction network GhostNet to extract multi-scale features of targets.Secondly，an improved Transformer is employed to construct the multi-scale feature fusion layer，which effectively integrates the multi-scale features of the target.Finally，the improved Transformer is combined with the cross-correlation operation of the Siamese network to integrate global and local information of features，achieving precise localization of the target.The experimental results conducted on the UAV123，GOT-10k，and LaSOT datasets demonstrate that the proposed algorithm exhibits excellent tracking performance with a speed of 87 frames per second，realizing real-time object tracking on both GPU and CPU platforms.

In response to the common problems of missed detections，false detections，and low detection accuracy in traditional anti-UAV detection methods，an improved YOLOv8 anti-UAV object detection algorithm，namely YOLO-DAP is proposed.Firstly，the large object detection layer (P5) is removed and the detection heads with size of 80×80，40×40 and 20×20 are replaced with new detection heads with size of 160×160，80×80 and 40×40，respectively，so as to improves the detection accuracy of small objects by the network. Secondly，DWR，an expandable residual attention module is introduced to improve Bottleneck block in C2f to improve the feature extraction ability of the network.Furthermore，ADown，a lightweight downsampling module，is introduced to better fuse feature maps of different scales.Finally，the PIoU loss function is used as the regression loss function，which makes the network have faster convergence speed and higher detection accuracy. Experiments on the public UAV data set TIB-Net show that the mAP of YOLO-DAP algorithm reaches 92.7%，which is 7.8 percentage points higher than that of the original YOLOv8n algorithm，and the number of parameters is reduced by 1.93×106，which also has obvious advantages compared with the other mainstream object detection algorithms，and the effectiveness and advancement of the algorithm is proved.

Aiming at the drawbacks of the traditional imaging algorithm for circular Synthetic Aperture Radar (SAR)，such as limited resolution，high sidelobes，and low imaging quality，a three-dimensional imaging algorithm for circular SAR combining Fractional Fourier Transform (FrFT) with Orthogonal Matching Pursuit (OMP) algorithm is proposed.First，the echo model of circular SAR is established.After selecting an excellent imaging algorithm，the energy concentration characteristics of Linear Frequency Modulation (LFM) signals in the fractional Fourier domain are utilized.The FrFT is employed for pulse compression to improve the resolution in the range direction.Then，based on the sparse properties of circular SAR signals，the OMP algorithm is used to estimate the azimuth and elevation angles of the signals in each range cell，achieving high-resolution imaging with a small amount of data.The simulation results show that: 1) The peak sidelobe ratio in the range dimension is 14% lower than that of the Iterative Adaptive Approach (IAA) algorithm and 13% lower than that of the OMP algorithm;and 2) In the angular dimension，it is 4% lower than that of the IAA algorithm.

Aiming at the problems of low matching accuracy and poor real-time performance of traditional local feature matching algorithms in complex scenes，an improved image matching algorithm based on descriptor fusion and AGAST algorithm is proposed.Firstly，the template image and the image to be matched are filtered by Gaussian filter.Secondly，the grid AGAST algorithm based on adaptive threshold is used to extract feature points，which makes the distribution of feature points more uniform.Thirdly，the BRIEF descriptor and BEBLID descriptor are fused for the first time，and a new descriptor named BRBLID is proposed. The FLANN algorithm is then used to perform a preliminary matching of the descriptors.Finally，the RANSAC algorithm is applied to eliminate the feature point pairs with large matching errors.Experimental results show that the improved algorithm performs well in image matching in complex scenes，and the matching accuracy and efficiency are greatly improved with higher reliability and robustness

The air-ground cooperative multi-agent system has gained widespread application in both military and civilian fields due to its excellent adaptability and functional complementarity. As a crucial component of the air-ground cooperative multi-agent system，the performance of the ground monitoring system directly affects the effectiveness of the whole system's cooperative functions.However，the research on the ground monitoring system of the air-ground cooperative multi-agent system is still in its infancy.Firstly，the structural design of the ground monitoring system is introduced，including the software architecture and development platform.Then，the key technologies involved in the development and implementation of ground monitoring system，such as communication protocols，task environment construction，and task planning，are discussed.Finally，the future development trends of the air-ground cooperative multi-agent ground monitoring system is predicted.

Hyperspectral anomaly detection is an unsupervised object detection algorithm designed to distinguish pixels that are significantly different from surrounding pixels.As an important means to achieve hyperspectral anomaly detection，local contrast usually adopts a fixed dual-window structure，which has average generalization ability and is easily affected by noise.To address these problems，a hyperspectral anomaly detection based on Taylor decomposition and adaptive window contrast is proposed.First，Taylor decomposition is used to decompose the spectral curve to further increase the difference between the background and the target.An adaptive window is used to ensure that the target is surrounded by the inner frame as much as possible to reduce the false alarm rate.The spectral angular distance is then used within the window to describe the differences in spectral curves.Finally，the contrast method is used to obtain the final anomaly detection results. Compared with seven advanced methods on four data sets，the results show that the proposed method has high detection accuracy and low false alarm rate.

A Complex-Valued Neural Network (CVNN) radar target recognition method based on echo in resonance region is proposed to solve the problem that the existing schemes has low utilization rate of echo in resonance region.CVNN can not only improve the problem of data loss in the process of poles feature extraction，and realize automatic feature extraction of features in the resonance echo，but also take into account the internal coupling relationship between the real part and the imaginary part of data.The artificially extracted pole features，the target features extracted by real-valued Convolutional Neural Network(CNN) and CVNN are input into KNN and SVM for comparison.Experimental results prove the effectiveness of the proposed method，and the recognition accuracy can still reach more than 98.5% under the condition of -5 dB signal-to-noise ratio.Compared with the existing schemes，the accuracy of target recognition is improved.

Aiming at the problem of poor detection performance of military objects at night under low-light and smoke occlusion，MLEC-YOLO model is proposed to be dedicated to the detection of military objects at night.Firstly，an enhanced low-frequency feature extraction network is constructed as the backbone network，by combining the multi-scale low-frequency feature extraction module and the dynamic feature fusion module，the multi-scale low-frequency feature extraction and key features perception are carried out respectively.Secondly，a deep path aggregation network is designed in the neck network to enhance the feature representation from the backbone network.Finally，four decoupling heads with different resolutions are adopted to accommodate military targets of different sizes in night scenes.Simulation results on the self-built dataset Nighttime~~Military and the public dataset BDD100K show that the proposed scheme significantly outperforms most of current mainstream object detection models in terms of detection accuracy and generalization ability.

Mobile platform cooperative detection radar system is a network composed of multiple radar nodes，which can significantly improve the target detection and anti-complex jamming ability.It is a hot research topic at present，but how to quantitatively evaluate its anti-jamming ability is a difficult issue.A method for evaluating the anti-jamming performance of mobile platform cooperative detection radar system is proposed.By analyzing the anti-jamming performance in different application scenarios，the anti-jamming capability enhancement degree is defined to quantitatively evaluate the anti-jamming performance of a single radar and the whole system.The evaluation steps include determining the reference scene，determining the scene to be evaluated，calculating the signal-to-jamming ratio and the improvement of anti-jamming performance，and verifying the effectiveness of this method through an example calculation.The proposed method provides an idea and a way for quantitatively evaluating the anti-jamming capability of the cooperative detection radar system of mobile platform.

In order to detect and identify unauthorized UAV more accurately，it is necessary to detect and identify the RF signal of UAV under the condition of low SNR.An identification method based on time-frequency spectrogram is proposed to solve the identification problem of UAV RF signal in low SNR.Firstly，the UAV signal is transformed into two-dimensional time-frequency spectrogram by short-time Fourier transform，which is used as the input of neural network.Then，an improved ResNet18 network is built，which introduces channel attention mechanism and spatial attention mechanism，and adopts regularization strategy and adaptive adjustment strategy to improve the identification accuracy of UAV RF signals in low SNR.The experimental results show that，when the SNR is -15 dB，the identification accuracy of proposed model for 16 types of UAV RF signals reaches 0.906 2，which is 0.074 0 higher than that of ResNet18 network，and its performance is better than that of network model methods such as EfficientNet，MobileNetv2 and GoogLeNet. It also shows improved performance under actual noise conditions，demonstrating better robustness to noise and anti-confusion capabilities.

In the task of signal sorting in the field of electronic countermeasures，in order to solve the problem that the traditional PRI-based method is difficult to sort in complex environments，a radiation source sorting method based on parallel denoising autoencoder is proposed.The method implements binary pre-processing of TOA sequences，constructs a parallel denoising autoencoder model，and realizes the synchronous sorting of various types of aliased signals.Experimental results show that the proposed method maintains high accuracy and good robustness under the influence of high error.Compared with the traditional algorithm，this method improves the sorting accuracy，and the accuracy of the algorithm can still be maintained above 90% when the pulse loss rate reaches 50%.

With rapid advancements in machine vision technology，various aviation safety agencies，research organizations，and leading companies are actively engaged in the development，test flights，and certification of aircraft Visual Landing System (VLS).Firstly，the basic working principle and operation concepts of VLS are investigated，and an airport runway detection model based on the YOLOv5 network architecture is proposed and simulated on the Landing Approach Runway Detection (LARD) dataset.Furthermore，all the VLS functions including runway detection are listed.On the basis of considering the uncertainty and generalization gap of machine learning model，the failure conditions of each function are identified and categorized according to the severity of their impacts，and the means of compliance used to verify the safety requirements related to the failure conditions is given.It is shown that the precision，recall，and mAP~~0.5 of the proposed model reaches 95%，96%，and 98% respectively，the results of functional hazard assessment provide ideas and references for the development of VLS，the functional definition of AI-based avionics system and the analysis of the failure conditions.

Traditional focusing evaluation functions suffer from poor auto-focusing stability and low sensitivity in the process of auto-focusing in UAV aerial photography due to the diversity of edge directions.In order to solve the problems，a TR focusing evaluation function based on the threshold is proposed.The function combines the advantage of the Tenengrad threshold function in reducing image noise interference and the advantage of the Roberts function in calculating gradient variations in multiple directions，and introduces a multi-directional edge gradient variation detection operator with the ability to extract gradient variations information in multiple directions.The experimental results show that，compared with the mainstream focusing evaluation functions such as SMD，Sobel，Roberts，Tenengrad，etc.，the TR focusing evaluation function has a significant advantage in two objective evaluation indexes of clarity ratio and sensitivity，with better stability and higher sensitivity，which is able to satisfy the auto-focusing needs of UAVs for accurately detecting ground targets in low-altitude environments.