The control problem of multi-UAV systems is studied based on differential game strategies.The goal is to minimize the performance metrics of each UAV.Under the open-loop Nash equilibrium condition，a finite-time cost function is introduced to transform the problem of multi-UAV system control into a linear quadratic differential game problem.Through coupled differential Riccati equations，optimal strategies under finite-time open-loop Nash equilibrium are derived，along with corresponding proofs.Subsequently，a method combining receding horizon with optimal control is employed，and the current state is taken as the initial condition for each time step，which achieves effective control of the multi-UAV system.Finally，simulation studies with five UAVs validated the effectiveness of the proposed method.

For global path planning of UAVs and the evasion of local dynamic obstacles during flight，a real-time dynamic obstacle avoidance algorithm that integrates the improved A* algorithm with the velocity obstacle method is designed.This algorithm introduces an adaptive node expansion step size strategy based on Euclidean Signed Distance Field(ESDF)map into the traditional A* algorithm to reduce redundant nodes in the path searching process.The initial path is smoothed by the B-spline curve.A velocity obstacle method in three-dimensional space is designed based on UAV dynamics model，achieving real-time obstacle avoidance in local dynamic environment.An attraction function is introduced between the local obstacle avoidance path and the global waypoints to guide the UAV to smoothly return to the global path after avoiding local obstacles，preventing it from falling into a locally optimal state.The simulation experiments show that the proposed algorithm can save the search time by 64.5% and reduce the number of exploration nodes by 62.2% for global path planning，and the UAV can effectively avoiding obstacles in dynamic environment.

To address the network connectivity issue in multi-UAV collaborative search，a collaborative search decision method incorporating network connectivity and Dung Beetle Optimization (DBO) is designed. Firstly，a novel probabilistic map updating strategy is established by simulating the target's decision state transition process using a Markov chain.To improve environmental certainty and reduce redundant searches，a certainty map and a digital pheromone map are introduced.Secondly，a real-time search objective function is constructed by combining search efficiency and the established minimum spanning tree network topology.The DBO enhanced with dynamic critical factors and adaptive t-distribution is then used to plan the optimal search paths for the UAVs.Finally，simulation results show that the proposed network topology improves target search efficiency by 17.84%.The proposed algorithm outperforms comparative algorithms in terms of coverage stability and convergence accuracy.

To solve the problems of serious missed detection and false detection caused by the large number of small target samples with a wide scale range in UAV aerial images，an improved small target detection algorithm named SD-YOLO is proposed based on YOLOv8s.Firstly，the C2f is reconstructed by using DCNv2 module，so that the model can effectively capture the fine-grained details of the target and adaptively adjust the sampling position of the convolution kernel，so as to accurately perform target positioning.Secondly，the SPD-Conv module is improved to enhance the model's ability of capturing local features，so that the model can retain more spatial information.Finally，a small target detection head is added，and the Dynamic head module is introduced to improve detection performance of the model in multi-scale scenes.Experimental results on VisDrone2019 dataset show that SD-YOLO has an mAP50 of 0.495，which is 0.1 higher than that of the original YOLOv8s network，and it is able to maintain a high frame rate，which significantly improves the detection performance of multi-scale small targets.

An improved YOLOv8n-based small target detection algorithm is proposed to address the problem of low detection accuracy caused by small，densely overlapping targets and diverse and complex scenes in drone aerial images.The improved deformable convolution is integrated with some C2f layers in the backbone network to enhance the network's adaptability to different target scale changes.The GC attention mechanism is introduced in the Neck part to enhance the network's feature extraction ability.In the Head part，a dynamic detection head，DyHead，is designed，which unifies the attention of three dimensions of scale，space，and task，to improve the detection effect of the detection head.The feature fusion structure and detection head of the original model are improved to effectively improve the network's detection accuracy for small targets and reduce the model parameter quantity.Experimental results on the VisDrone dataset show that:1) The mAP50 of the improved model has increased by 6.0 percentage points compared with that of the base model，and the frame rate when deploying on hardware is 20 frames per second;and 2) The detection effect of this model is significantly improved，and the detection speed meets the real-time requirements for small target detection in UAV aerial images.

Semantic segmentation of remote sensing images is widely applied in such fields as environmental change monitoring and automotive driving assistance.Remote sensing images exhibit great intra-class variability and small inter-class differences at the semantic object level，which limits the accuracy of segmentation models and consumes computational resources.To address the problems，this paper proposes a lightweight semantic segmentation method for remote sensing images based on Transformer and depth-wise separable convolution.Firstly，a weight-adaptive multi-head self-attention mechanism is introduced to model the long-range pixel associations in a global scale，capturing rich contextual information.Secondly，stacked layers of Depth-wise Separable Convolutions (DSCs) are constructed to reduce the loss of spatial detail information with low computational complexity.Additionally，a Feature Aggregation Module (FAM) is designed by using a Linear Attention (LA) mechanism to merge global scene information and spatial detail information.The tests on datasets of Vaihingen and Potsdam show that the proposed method achieves an overall segmentation accuracy of 92.6% and 92.1% respectively with GFLOPs of only 11.5.It not only effectively enhances segmentation precision，but also significantly reduces computational complexity.

In highly dynamic and rapidly changing battlefield environments，satellite navigation systems are often in a denial state，and the accuracy of the navigation system for heterogeneous missile flight clusters，originating from different sources and equipped with different navigation sensors，is severely affected，resulting in limited flight range.To address this challenge，a novel cooperative navigation system for heterogeneous missile clusters is designed.This system relies solely on the ranging information from inter-missile data links，and employs rank-loss constrained cascaded filtering technology to achieve real-time correction of navigation errors of each node within the heterogeneous missile cluster.Digital simulation verification and UAV flight test results indicate that the proposed method can effectively improve the navigation accuracy and adaptability of heterogeneous clustered missiles in complex battlefield environments，which has significant reference value for the design of navigation systems in future formations of networked weapons.

Sparse aperture ISAR imaging involves reconstructing high-resolution ISAR images from incomplete echo data.Currently，sparse recovery methods are mainly classified into two kinds of model-driven imaging methods and data-driven imaging methods.This paper first introduces the basic principle of ISAR imaging and the signal model for sparse aperture，then discusses three classes of model-driven compressive sensing methods，namely convex relaxation optimization algorithms，non-convex optimization algorithms，and greedy algorithms，and compares and analyzes the advantages and disadvantages of each method. Subsequently，deep learning-based sparse imaging methods for ISAR are introduced，focusing on neural network based learning and deep unfolding network based learning，with evaluations of their efficacy in ISAR sparse imaging applications.Finally，the content of this paper is summarized and the development trends of sparse aperture ISAR imaging are given.

In the field of radar imagery，the moving target detection task has been receiving a lot of attention. In recent years，with the development of electronic technology and radar technology，Video Synthetic Aperture Radar (Video-SAR) technology brings SAR image data with higher frame rate.At the same time，the excellent performance of deep learning in the target detection task also makes many scholars apply it to the Video SAR moving target detection task.The Video-SAR moving target detection techniques have begun to shift from the traditional Ground Moving Target Indication (GMTI) approach to the deep learning approach.However，deep learning-based Video-SAR moving target detection methods are still in the research stage，and there are no review articles on the Video-SAR moving target detection task. Therefore，this paper analyses and organizes the existing methods at home and abroad on the basis of moving target detection methods，summarizes the unsolved problems and difficulties faced by the Video-SAR moving target detection task in combination with the existing research，and puts forward research suggestions for its future development based on the development of existing technologies.

To address the issue of heterogeneous communication among multiple modules during the digital simulation phase of civil avionic system design，and improve the efficiency of collaborative simulation across multiple modules and systems，a fully digital simulation platform based on Data Distributed Service (DDS) is constructed.This platform includes three subsystems，namely，console，configurator and data monitor.Automatic configuration of the simulation environment and automatic generation of the simulation framework are achieved via an aviation Interface Control Document (ICD).A standardized message hierarchy is devised for various aviation bus protocols，and a generic communication interface is established using DDS middleware to facilitate multi-protocol data exchange functionality.This ensures dependable communication among multiple simulation modules across heterogeneous platforms.Experimental results demonstrate that the maximum latency of the platform is controlled within 1 ms，with a jitter of less than 200 s，and a data transmission rate of up to 800 Mibit/s，which effectively meets the data transmission performance requirement of avionics simulation.

As a landing mode of fixed wing aircraft，visual landing will exist for a long time.Based on the visual landing principle，the internal mechanism of the light array optical landing aid system is analyzed，and the light array arrangement scheme is analyzed considering the aircraft landing characteristics，aircraft carrier characteristics and the influence of other factors.According to the landing glide path geometry，the visual landing point position correction algorithm under two kinds of stable modes is proposed，and the light array unit control logic is optimized to improve the landing guidance accuracy.Finally，the designed light array optical landing aid system is simulated and verified by experiments.The result shows that the system realizes the unity of pilot's line of sight in “watching lights” and “centering”，and can meet the requirements of guidance accuracy and landing safety，which is an important research direction of visual landing guidance method.

In order to address the issues of low brightness and halo in bright areas in images processed by the dark channel prior algorithm，an improved dehazing algorithm based on improved superpixel segmentation is proposed.Firstly，the traditional dark channel window is replaced by a superpixel segmentation window，and the image transmittance is preliminarily improved by combining adaptive superpixel blocks with dual window fitting.Then，the transmittance of the bright area in the image is exponentially corrected separately，and then weighted and fused with the initial transmittance to further improve the transmittance of the bright area.Trapezoidal low-pass filtering is used to smooth and denoise the improved dark channel，and quadtree subdivision is performed on the smoothed dark channel to obtain more accurate atmospheric light.The experimental results show that the algorithm effectively improves the defects of the dark channel prior algorithm in restoring bright areas of images，repairs halo artifacts，and achieves good dehazing effects.

To address the issue of Unmanned Aerial Vehicles (UAVs) flight safety in urban spaces，the “flight in separated air corridors” mode is put forward.The safety limit for each type of UAV is defined as the vertical width required for the conflicting UAVs to avoid each other within these corridors，and it serves as the basis of corridor demarcation.The core of the algorithm is Velocity Obstacles (VO).To eliminate the overly conservative assumption of the classic VO circular domain，which causes overly large safety margins，this paper proposes a dynamic elliptical protection domain based on the relative positions of the conflicting UAVs and its adaptive algorithm. To prevent overly large safety margins for long corridors and avoid deviations from the original route after obstacle avoidance，this paper designs a fast and smooth trajectory restoration control method that also eliminates velocity oscillations.The simulation experiments under two preset scenarios demonstrate that the elliptical domain not only maintains the same safety interval as the circular domain，but also requires smaller safety margins on the corridors' boundary.This contributes to real-time obstacle avoidance capability and a reduced diversion distance.

In order to measure high-speed targets，a direct detecting lidar based on microwave photonics method，called microwave lidar，is designed.The basic principle is to use the phase modulator to load the radio frequency (microwave) signal onto the outgoing laser of the lidar，then the Fabry-Perot interferometer and the photodetector work together to demodulate the radio frequency signal from the target echo signal received by the lidar，and thus the target information is obtained by measuring the radio frequency signal.The microwave lidar retains the characteristics of high spatial resolution of the lidar，and also exerts the advantages of high-speed measurement range and anti-atmospheric turbulence of the microwave radar，thereby providing a new approach for the detection of high-speed targets.The software radio technology is used to generate，receive and process radio frequency (microwave) signals，and the speed measurement of rotating hard targets and aerosol simulated targets (such as high-speed missiles and their shock fields) is made.The experimental results proves that the lidar can measure high-speed targets (within Mach 4.4) and their shock fields.

Optical axis stability is an important index to determine the detection capability of infrared system in the airborne random vibration environment.In the design stage，a continuous zooming infrared system is taken as the research object.Firstly，the moving lens is identified as the optical axis sensitive element.An accurate finite element model is established for the linear motion mechanism of the lens.Through the integrated optical and mechanical analysis method，the optimization design of the optical and mechanical structure of the continuous zooming infrared system is completed.The finite element modeling method of linear motion mechanism of lens and the optimization design process of optical axis stability proposed in this paper have important reference significance for the subsequent design of random vibration optical axis stability of airborne photoelectric imaging system under random vibration.

To achieve autonomous UAV flight path planning for policing tasks，an environment model is constructed for police UAV based on real urban police emergency response scenarios.A fitness function is designed according to the costs of UAV's path length，flight altitude and obstacle danger.Logistic chaotic map is used to initialize the population in the Aquila Optimizer(AO)，and the Butterfly Optimization Strategy(BOS) is incorporated during the first phase.In the simulation experiments，the generated paths are smoothed by using cubic spline interpolation to make them more in line with the trajectory of the UAV.The results show that this method can meet the requirements for path planning of police UAVs in rapid dispatch，and the overall effects are satisfactory.

In weapon range testing，the position of the detonation point provides crucial data support for fuse performance evaluation.To rapidly and accurately identify detonation point regions in images，a detonation point image recognition algorithm is proposed based on YOLOv5-FireCAM network.The method improves the original YOLOv5s network by introducing a lightweight Fire module and using a new C3-Fire network to replace original backbone network.A CAM color attention module is added at the end of the backbone network to enhance detonation point image feature extraction.In the section of the detection head，an Anchor-Free(AF) structure is used to reduce network complexity and computational load while increasing detection accuracy.During model training，the Softmax loss with Hard Negative Mining is adopted as the classification loss function，and the smooth L1 loss function is applied for network bounding box regression in view of the AF structure.Experimental results show that the proposed detection network algorithm achieves an mAP@0.5 of 88.4% on the detonation point dataset，with 28.1% improvement in detection speed and 32.8% reduction in network parameters in comparison with those of original network.This method maintains fast detonation point region identification with high accuracy while reducing network parameters，which is faster and more accurate than the original algorithm.