To settle the problem of low tracking precision caused by fixed Transition Probability Matrix (TPM) of standard Interacting Multiple Model (IMM) algorithm, an Adaptive IMM (AIMM) algorithm of TPM is introduced. Based on the probability change information of adjacent time and the posterior probability information of different models at current time, an element correction function of Markov matrix is introduced to adjust each element in the transition probability matrix in real time. Besides, Cubature Kalman Filter (CKF) is introduced into IMM algorithm. The simulation experiments are set to testify the availability of the proposed algorithm. The results show that in comparison with the traditional IMM-CKF algorithm and other adaptive IMM algorithms, the proposed AIMM-CKF algorithm effectively improves model matching probability and model switching rate, which exhibits better tracking effects.

Aiming at the problems of external interferences and actuator failure faced by the quadrotor UAV formation, an Adaptive Barrier Function Fast Terminal Sliding Mode Control (ABFFTSMC) is proposed. Firstly, a distributed formation control method is designed, which consists of an undirected fixed topology for communication between quadrotor UAVs, thus communication is realized only by using the information of neighboring UAVs, reducing the demand for global data and the pressure on communication bandwidth. A fast terminal sliding mode controller with adaptive barrier function is designed. The variable gain is used to compensate for interference and faults, and the tracking errors converge to zero in a finite time. The simulation results show that this algorithm can greatly improve the fault tolerance performance of formation of multiple quadrotor UAVs under the conditions of external interferences and actuator failure.

In order to improve the positioning accuracy of the positioning systems based on One-Dimensional Angle of Arrival (1-D AOA) measurement, this paper proposes an optimal observation station deployment method based on the Improved Chicken Swarm Optimization (ICSO) algorithm. Firstly, the trace minimum of the Cramr-Rao Lower Bound (CRLB) is adopted as the optimization criterion to establish the optimal solution model. Secondly, in order to solve the high-dimensional variable problem caused by a large number of observation stations, the updating strategy of the Chicken Swarm Optimization (CSO) is improved. Finally, ICSO is used to optimize the locations of observation stations and the orientation of the linear array. The simulation results show that:1) ICSO has higher convergence speed and higher positioning accuracy in optimal station deployment solving; and 2) The proposed optimal station deployment method can significantly improve the positioning accuracy under different numbers of observation stations, and the dependence of positioning accuracy on the number of observation stations can be reduced by optimizing the location of observation stations in engineering practice.

Aiming at the problems of long shooting distance, small target, high density and mutual occlusion of objects in aerial images of Unmanned Aerial Vehicles(UAVs), an improved YOLOv8s algorithm called DD-YOLO is introduced, which combines deformable depth-wise convolution and multiple attention mechanisms. The algorithm integrates depth-wise convolution to simplify the network model, and proposes deformable depth-wise convolution to optimize C2f module and enhance the feature extraction ability of backbone network. The SE and MHSA attention mechanisms are introduced to transform the structure of SPPF to make it take into account the extraction of local and global features. Quadruple downsampling branches are added to the neck network to alleviate the lack of receptive field for small targets, optimize target localization, and strengthen the focus on small targets. Experiments show that the improved model achieves an mAP@50 of 43. 9% and a mAP@50∶95 of 26. 7% on the VisDrone-DET2019 dataset, which is 5. 1 and 3. 6 percentage points higher than that of YOLOv8s, respectively, with a 13. 2% reduction in parameter quantity and 12. 6% reduction in model size which is of great significance for the realization of UAV small target detection.

Focusing on the multi-UAV formation collision avoidance problem, a prescribed-time distributed control method based on scaling potential function is proposed. Firstly, a non-jump time-varying function is designed to improve the prescribed-time stability criterion, which achieves fast and smooth convergence of systems. Secondly, a prescribed-time distributed observer is designed to compensate the composite disturbance without relying on the system disturbance and its derivatives. Thirdly, based on the above observer, the leading UAV's obstacle avoidance control strategy is further constructed by designing the scaling potential function, and a follower UAV collision avoidance control method is proposed based on the error transformation function, so that the UAVs can maintain the original formation while avoiding obstacles. Finally, the feasibility of the proposed method is proved by theoretical analysis and numerical simulation.

To tackle the challenge of multi-sensor information fusion in the absence of prior knowledge, this paper introduces a multi-sensor information evaluation fusion algorithm that incorporates information entropy and the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) from the perspective of evaluating the quality of information obtained by sensors in order to obtain results closer to the true value. Firstly, information entropy is used to evaluate the consistency of information obtained by all the sensors at each sampling moment and to weight the sampling moments. TOPSIS is used to evaluate the quality of information obtained by all the sensors and to weight each sensor according to quality ranking. Then, according to the weight of sampling moments and the weight of sensors, a temporary fusion result is obtained and then iterated so that it will converge to the true value. Finally, through a series of simulation experiments, it is demonstrated that the proposed information evaluation fusion algorithm significantly outperforms the existing similar algorithms in terms of stability and accuracy.

Simultaneous Localization and Mapping (SLAM) is one of the important research directions in the fields of robotics, autonomous driving, and other fields nowadays. In response to current situation where monocular vision odometry technology is unable to extract sufficient feature points in weak feature environments with few feature points in external environments, a point-line feature fusion algorithm Based on Point-Line Feature fusion Visual Inertial Navigation Systems (BPLF-VINS) is proposed, which takes the mainstream algorithm Monocular Visual Inertial Navigation Systems (VINS-Mono) as the basic framework. An adaptive point-line feature extraction strategy based on different external environments is adopted, which can maximize the role of line features and avoid system resource waste. In addition, the proposed line segment suppression strategy can ensure the quality of line features, which improves positioning accuracy while maintaining system operation efficiency. The improved LSD algorithm has increased efficiency by 50. 07%. In addition, in order to address the issue of long line segments being mistakenly detected as short line segments, an improved line segment merging strategy is proposed to further improve the accuracy of line segment matching. Through a series of testing experiments on authoritative dataset EuRoC, it is shown that the Root Mean Square Error (RMSE) of the improved algorithm is decreased by an average of 16. 03% and 19. 69% respectively compared with that of PL-VINS and VINS-Mono. The experimental results verify the effectiveness of the algorithm.

In order to extend GAN to multi-tasking mode and construct a high-performance model, this paper combines Reinforcement Learning (RL) agents with GAN to construct a multi-tasking image generation model, RL-GAN. The model performance is improved by replacing the RL agent training algorithm, setting a more reasonable AC network loss function, and replacing the network structure. The experimental results show that:1) The generated results of the model in two multi-tasking image restoration experiments meet visual requirements; 2) Compared with multi-GAN stacking, a mainstream method in current multi-tasking modes, the RL-GAN model has faster convergence and image processing speed, higher output quality, and the accuracy and efficiency of the model are also better after introducing RL agents; and 3) The optimized model significantly improves its multi-tasking processing ability.

Aiming at the defects of limited receptive field in the deblurring of remote sensing images based on convolutional neural network, which leads to the loss of details and incomplete deblurring of images in the recovery process, a multi-scale remote sensing image deblurring algorithm integrating self-attention is proposed. A multi-input, multi-output U-Net is utilized to simulate a single U-Net into a multi-level joint multi-scale convolution operation to achieve effective extraction of features. A Transformer-based multi-head self-attention module is proposed, which enhances the network's spatial feature extraction and global information capture capabilities by embedding it in the middle position of encoder and decoder. A multi-scale edge loss function is introduced to improve the restoration of image edge details. A fuzzy remote sensing image dataset is constructed for the experiment, and quantitative and qualitative analysis of the experimental results shows that the proposed algorithm is superior to the contrast algorithms. To demonstrate the generalization ability of the algorithm, it is validated on the public dataset GOPRO. The results show that the algorithm is of certain practical significance for effective processing of blurred remote sensing images.

As a passive navigation mode, the scene matching navigation of UAVs has been extensively studied. Feature point extracting and matching are important components in scene matching navigation of UAVs. Traditional feature point extracting and matching algorithms do not provide global negative feedback of the results, resulting in low accuracy in heterogeneous image feature matching. To solve the problem, this paper proposes a UAV scene matching navigation method based on deep neural network feature matching. In this method, the deep neural network of SuperPoint and algorithm of LightGlue are introduced and improved for feature point extracting and feature matching, and the accuracy and stability of feature matching are improved. In order to solve the problem of huge difference between pixels in heterogeneous images, an image grayscale conversion algorithm is designed, which effectively reduces the influence of pixel differences on the matching results in the image matching process. Finally, the simulation experiment results show that, compared with traditional ORB algorithm, the deep learning algorithm can solve the feature matching problem of UAVs in complex environments more effectively.

Modern information-based war is a kind of confrontation between systems. It is necessary to build a corresponding operation system through equipment coordination. The overall efficiency can be doubled by the collaboration of multiple UAVs. How to reasonably arrange the task execution order of UAVs based on the position of multiple UAVs has become the core issue of UAVs mission planning. A multi-UAV mission planning method based on feature information mining is designed. Firstly, a data mining algorithm is proposed. The algorithm has a new data structure and can store utility, time, location and other information, which is suitable for subsequent research of multi-UAV mission planning methods. Then, the proposed algorithm is used to mine the simulation data of multi-UAV mission scenarios, and the optimal multi-UAV mission planning method under different threat scenarios is obtained. Finally, the simulation results are optimized according to the obtained model, and comparison is made on the efficiency before and after optimization.

To address the issues of low registration efficiency, prone to false matching, and difficulty in processing complex point cloud data in the data registration stage of point clouds obtained through lidar scanning, a point cloud registration method is proposed based on the optimization of the sampling rate and the improvement of descriptor extracting. Firstly, the optimal sampling rate for extracting key points is sought based on the influence of the sampling rate on key point extracting, and the voxel filtering method is improved to achieve optimal down-sampling. This ensures the registration accuracy while achieving data streamlining. Secondly, considering the inefficiency of constructing descriptors for global data for corresponding point matching, the key point extraction strategy is utilized to achieve a simplified construction of descriptors and improve the efficiency of corresponding point matching. Then, dual checks are introduced to reduce the probability of false corresponding point matching. Finally, comparative registration experiments are conducted using the actual measurement data of lidar and open-source point cloud datasets. The experimental results indicate that the proposed registration method has higher registration speed while keeping the matching accuracy of point clouds.

To solve the problems of low detection rate and high false alarm rate caused by clutter interference in the detection of inshore ship targets in Synthetic Aperture Radar (SAR) images, an SAR image ship detection algorithm for complex inshore scenarios is proposed. the target feature extraction ability of the feature extraction network is enhanced by designing C2f-EMBC and BasicStage based on YOLOv8n, which captures the geometric feature information of the ships, and makes the network pay more attention to detailed features. Meanwhile, a Global Receptive Field Spatial Pyramid Pooling Fast (GRF-SPPF) algorithm is proposed, which combines important information of the feature layer with background information of global receptive field. In the feature fusion structure, AKConv and Slim-Neck are used to design the AVSFPN neck structure for feature fusion at different levels. In addition, the WIoU loss function is used to improve the convergence rate and generalization of the model. The testing experiments are conducted on the BBox-SSDD and HRSID datasets. The results show that: the improved algorithm has an mAP0. 5 of 96. 4% and 77. 4% respectively on the BBox-SSDD and HRSID inshore test set, which is increased by 4. 8 and 4 percentage points respectively in comparison with the original YOLOv8n. The results prove the effectiveness of the proposed method in improving the accuracy of ship target detection in SAR images in inshore scenarios.

In order to solve the problem of low detection accuracy caused by small size and dense distribution of targets, complex detection scene and target occlusion from the viewpoint of UAV aerial photography, an SSD-based small target detection algorithm is proposed. Firstly, the SE-ResNet50 network is used to replace VGG16 as the backbone network, and the model pays more attention to useful channel information by learning the adaptive channel weights. Then, the anchor frame parameters of the shallow-layer network are modified to improve the performance of small target detection. In the shallow-layer network, two combinations of SE+SAM and CBAM are adopted to pay attention to the images from the two dimensions of channel and space. Finally, EIoU is used instead of traditional IoU, to calculate the intersection over union ratio, and the FocalL1 loss function is adopted. Then, the EIoU and FocalL1 loss functions are integrated to obtain final Focal-EIoU. VisDrone2019 dataset is used for verification. In comparison with that of traditional SSD algorithm, the precision, recall and mAP is increased by 11. 7, 11. 2 and 9. 8 percentage points respectively, and FPS is increased by 5 frames per second, which verifies the effectiveness of the algorithm in small target detection.

In order to solve the problem that the Radar Cross Section (RCS) of the front photoelectric pod of a UAV is too large under the condition of maintaining high optical transmittance, a multi-layer, mesh-type, meta-surface is proposed. According to the equivalent theoretical circuit model of the cross-shaped metal mesh, the influence of electrical impedance on the electromagnetic shielding effect is analyzed, and the cross-shaped mesh is extended to a three-layer, mesh-type, meta-surface structure, and the equivalent circuit model is established. The simulation results show that the mesh-type meta-surface has a smaller RCS than the ordinary periodic mesh. Moreover, this multi-layer structure is helpful for integrating the RCS reduction structure into the front photoelectric pod of the UAV. Finally, the RCS test of the sample verifies the effectiveness of the proposed multi-layer, mesh-type, meta-surface structure, and achieves a positive-incidence RCS reduction of 10 dB in the frequency band of 12~14 GHz under the condition of high optical transmittance.

In order to solve the problem that it is difficult to correct the complex aberration introduced by conformal light windows in large aperture and large field-of-view systems, a new method is proposed for the design of airborne optical systems. This method adopts the inner surface of light windows and the discrete reflective mirror for aberration correction, balances the aberration through sub-aperture correction strategy, and gives a control algorithm that matches with the discrete reflective mirror to solve the problem of wave surface inconsistency between sub-apertures, which has the advantages of high degree of freedom and simple structure. The simulation experiments verify that the method can correct the wave aberration RMS of the central field of view from 0. 62 to 0. 06(=950 nm) at 120 mm aperture and 0°～30° scanning angles. The research shows that the discrete reflective mirror has good performance in dynamic aberration compensation, and the proposed correction method has wide application potential in the design of conformal optical systems.

As for the heterogeneous multi-agent system composed of multiple underactuated Unmanned Surface Vehicles (USVs), fixed-time distributed formation cooperative tracking control with dynamic uncertainties of heterogeneous USV system is studied. Firstly, the model coordinates of the controlled heterogeneous USV system are transformed and converted into a fully-driven, second-order heterogeneous multi-agent system. Then, for the uncertain model dynamics of each heterogeneous multi-agent system, a fixed-time dynamics observer is designed. On the basic of backstepping method and the virtual leader model, a fixed-time distributed cooperative tracking control protocol based on fixed-time uncertainty dynamics estimation is designed for the follower controller of each heterogeneous USV system. By constructing a suitable Lyapunov function, the convergence of the heterogeneous USV formation cooperative tracking error within a fixed time period is analyzed. Finally, the numerical simulation of formation cooperative tracking control is conducted in the cases of steady formation and time-varying formation respectively, which verifies the effectiveness of the designed fixed-time formation control algorithm.