Due to the limitations of small targets in remote sensing images, such as background noise, poor information, and so on, the results of commonly used detection algorithms in small target detection is not satisfactory. To improve the accuracy of detection results, we develop an improved algorithm based on YOLOv8, called LAR-YOLOv8. First, in the feature extraction network, the local module is enhanced by using the dual-branch architecture attention mechanism, while the vision transformer block is used to maximize the representation of the feature map. Second, an attention-guided bidirectional feature pyramid network is designed to generate more discriminative information by efficiently extracting feature from the shallow network through a dynamic sparse attention mechanism, and adding top–down paths to guide the subsequent network modules for feature fusion. Finally, the RIOU loss function is proposed to avoid the failure of the loss function and improve the shape consistency between the predicted and ground-truth box. Experimental results on NWPU VHR-10, RSOD, and CARPK datasets verify that LAR-YOLOv8 achieves satisfactory results in terms of mAP (small), mAP, model parameters, and FPS, and can prove that our modifications made to the original YOLOv8 model are effective.

Recent advances in the fields of driverless cars, intelligent robots and remote-sensing measurement have shown that the use of LiDAR fused with cameras can provide more comprehensive and reliable sensing of surroundings. However, since it is difficult to extract features from sparse LiDAR data to create 3D–2D correspondences, finding a method for accurate external calibration of all types of LiDAR with cameras has become a research hotspot. To solve this problem, this paper proposes a method to directly obtain the 3D–2D correspondences of LiDAR–camera systems to complete accurate calibration. In this method, a laser detector card is used as an auxiliary tool to directly obtain the correspondences between laser spots and image pixels, thus solving the problem of difficulty in extracting features from sparse LiDAR data. In addition, a two-stage framework from coarse to fine is designed in this paper, which not only can solve the perspective-n-point problem with observation errors, but also requires only four LiDAR data points and the corresponding pixel information for more accurate external calibration. Finally, extensive simulations and experimental results show that the effectiveness and accuracy of our method are better than existing methods.

LiDAR-guided stereo matching for high-precision disparity estimation is a very promising task in photogrammetry and remote sensing. Unfortunately, existing methods suffer from the problem that it is difficult to automatically obtain appropriate stereo matching model parameters to ensure satisfactory results. To solve it, this letter proposes a LiDAR-guided stereo matching framework using Bayesian optimization with Gaussian process regression, which aims to automatically infer the stereo matching model parameters by LiDAR data. First, local matching model based on the belief propagation algorithm is designed. Second, the objective function is constructed by minimizing the difference between the local matching results and the LiDAR data. Third, Bayesian optimization with Gaussian process regression is applied to minimize this objective function to infer the model parameters. Finally, experimental results on the GaoFen-7 and UAV Stereo datasets show that the proposed method can effectively infer suitable model parameters from LiDAR data, and our method outperforms the state-of-the-art methods.

Camera pose estimation plays a crucial role in computer vision, which is widely used in augmented reality, robotics and autonomous driving. However, previous studies have neglected the presence of outliers in measurements, so that even a small percentage of outliers will significantly degrade precision. In order to deal with outliers, this paper proposes using a graduated non‐convexity (GNC) method to suppress outliers in robust camera pose estimation, which serves as the core of GNCPnP. The authors first reformulate the camera pose estimation problem using a non‐convex cost, which is less affected by outliers. Then, to apply a non‐minimum solver to solve the reformulated problem, the authors use the Black‐Rangarajan duality theory to transform it. Finally, to address the dependence of non‐convex optimisation on initial values, the GNC method was customised according to the truncated least squares cost. The results of simulation and real experiments show that GNCPnP can effectively handle the interference of outliers and achieve higher accuracy compared to existing state‐of‐the‐art algorithms. In particular, the camera pose estimation accuracy of GNCPnP in the case of a low percentage of outliers is almost comparable to that of the state‐of‐the‐art algorithm in the case of no outliers.

When the Risley prism pair is used for target tracking, the nonlinear relationship between beam deflection and prism rotation makes tracking performance highly dependent on precise and stable motor control over a wide speed range. Although the brushless DC motor serves as the preferred drive source, its inherent commutation torque ripples directly induce beam pointing jitter, severely degrading overall tracking accuracy and stability. To address these issues, this paper proposes a three-vector-based model predictive direct speed control method. This approach establishes a direct speed-to-torque control channel by generating reference active power through dynamic equations, eliminating the need for fitting a constant flux linkage and parameter tuning. Simultaneously, combined with three-vector optimization and seven-segment modulation strategies, it achieves a dynamic balance between high-frequency, instantaneous electromagnetic power fine-tuning and inherent mechanical inertia of the rotor. Simulation results demonstrate that the proposed method exhibits superior speed stability compared to the conventional double-vector-based model predictive power control method and maintains high-precision dynamic tracking over a wide speed range. Ultimately, it leads to an average reduction of over 60% in the time-weighted absolute tracking error integral under various target trajectories, providing an effective solution for drive control of target tracking in Risley prism systems.