Unmanned aerial vehicle (UAV) autonomous tracking and landing is playing an increasingly important role in military and civil applications. In particular, machine learning has been successfully introduced to robotics-related tasks. A novel UAV autonomous tracking and landing approach based on a deep reinforcement learning strategy is presented in this paper, with the aim of dealing with the UAV motion control problem in an unpredictable and harsh environment. Instead of building a prior model and inferring the landing actions based on heuristic rules, a model-free method based on a partially observable Markov decision process (POMDP) is proposed. In the POMDP model, the UAV automatically learns the landing maneuver by an end-to-end neural network, which combines the Deep Deterministic Policy Gradients (DDPG) algorithm and heuristic rules. A Modular Open Robots Simulation Engine (MORSE)-based reinforcement learning framework is designed and validated with a continuous UAV tracking and landing task on a randomly moving platform in high sensor noise and intermittent measurements. The simulation results show that when the moving platform is moving in different trajectories, the average landing success rate of the proposed algorithm is about 10% higher than that of the Proportional-Integral-Derivative (PID) method. As an indirect result, a state-of-the-art deep reinforcement learning-based UAV control method is validated, where the UAV can learn the optimal strategy of a continuously autonomous landing and perform properly in a simulation environment.

The detection of dim and small space-targets is crucial in space situational awareness missions; however, low signal-to-noise ratio (SNR) targets and complex backgrounds pose significant challenges to such detection. This paper proposes a space-target detection framework comprising a space-target detection network and a k-means clustering target centroid positioning method. The space-target detection network performs a three-dimensional convolution of an input star image sequence to learn the motion features of the target, reduces the interference of noise using a soft thresholding module, and outputs the target detection result after positioning via the offsetting branch. The k-means centroid positioning method enables further high-precision subpixel-level centroid positioning of the detection network output. Experiments were conducted using simulated data containing various dim and small space-targets, multiple noises, and complex backgrounds; semi-real data with simulated space-targets added to the real star image; and fully real data. Experiments on the simulated data demonstrate the superior detection performance of the proposed method for multiple SNR conditions (particularly with very low false alarm rates), robustness regarding targets of varying numbers and speeds, and complex backgrounds (such as those containing stray light and slow motion). Experiments performed with semi-real and real data both demonstrate the excellent detection performance of the proposed method and its generalization capability.

The evolution of satellite surveillance technology, coupled with advanced on-board intelligent systems and improved attitude maneuver capabilities, has positioned mission scheduling and execution as a prominent and active research area in recent years. With the urgent need for mission scheduling and execution to evolve from the static ground target to the time-sensitive moving target, the existing satellite often yields unsatisfactory results when managed through conventional manual methods for continuously tracking a time-sensitive moving target. In response to this challenge, the paper introduces a reinforcement learning driven method enabling the satellite to schedule and execute the tracking of time-sensitive moving target autonomously. In the paper, we conduct a meticulous analysis of the target tracking process, which we decompose into three distinct sub-phases and address it utilizing attitude control, elucidating the integration of our model with satellite attitude dynamics and kinematics. Besides, we propose Flight Trajectory Prediction Module to predict the future trajectory position and minimize the position error due to the communication latency. With the dual aim of expeditiously tracking the target and concurrently enhancing payload stability, a well-designed reward function, which is applied to various reinforcement learning (RL) algorithms, effectively governs satellite attitude control to achieve accurate and stable tracking of moving target. In addition, through the eight distinct tracking scenarios simulated in our proposed Satellite Target Tracking Environment (STTE), our extensive experiments and ablation studies consistently underscore the advantages and efficacy of our proposed method. This work serves as a foundational cornerstone for future research endeavors focused on tracking multiple time-sensitive moving targets within extensive and complex intelligent satellite constellation.

Infrared (IR) small target detection has been widely used in military and civil fields. However, due to the insufficient feature integration capabilities of existing methods, effectively separating strong background clutter and targets in complex scenes remains difficult. To address this issue, we propose a two-stage feature complementary improved tensor low-rank sparse decomposition (TSFC-TLRSD) method. The detection process is divided into two stages: tensor initialization and tensor decomposition, effectively integrating local and non-local features. In the tensor initialization stage, inspired by the local saliency of the target and the local consistency of the background, we design a three-layer directional filtering (TLDF) operator for preliminary clutter suppression and target enhancement. Then, to promote the complementary advantages of local and non-local features, we refer to the TLDF and the original image to provide a targeted initialization strategy for the tensor low-rank sparse decomposition (TLRSD) model. In the tensor decomposition stage, we develop a robust partial sum of the tubal nuclear norm (RPSTNN) as a non-convex approximation of tensor rank, which can adaptively adjust the singular value distribution, thus adapting to diversity scenes. Meanwhile, we finely adjust the balance between low-rank and sparse components in the model-solving process through a nonlinear reweighting strategy, accelerating the optimization convergence speed and improving the model's background recovery ability. Extensive experiments on five practical datasets demonstrate that the proposed method is more effective and robust compared to ten state-of-the-art approaches.

Vessel trajectory prediction plays a crucial role in ensuring the safety and efficiency of maritime transportation. This study proposes an innovative sequence-to-sequence model, called the Vessel Influence Long Short-Term Memory (VI-LSTM), which introduces a novel Vessel Influence Map (VIM) to quantitatively model the dynamic effects of surrounding vessels. To enhance reliability, VI-LSTM incorporates Gaussian distribution predictions combined with Monte Carlo dropout techniques to estimate prediction uncertainty. Additionally, a temporally weighted hybrid loss function is designed to balance prediction accuracy and uncertainty. Furthermore, this study systematically categorizes and models factors influencing vessel trajectory prediction. Experimental results demonstrate that VI-LSTM achieves a mean distance error of 330.66 m on the standard test set and 480.30 m on an unseen subject test set, outperforming other comparative models, particularly in complex navigation scenarios and high-density maritime environments. These innovations significantly improve the accuracy and generalizability of vessel trajectory predictions, leading to enhanced safety, increased efficiency, and more effective collision avoidance in maritime navigation.

Current lidar-inertial SLAM algorithms mainly rely on the geometric features of the lidar for point cloud alignment. The issue of incorrect feature association arises because the matching process is susceptible to influences such as dynamic objects, occlusion, and environmental changes. To address this issue, we present a lidar-inertial SLAM system based on the LIO-SAM framework, combining semantic and geometric constraints for association optimization and keyframe selection. Specifically, we mitigate the impact of erroneous matching points on pose estimation by comparing the consistency of normal vectors in the surrounding region. Additionally, we incorporate semantic information to establish semantic constraints, further enhancing matching accuracy. Furthermore, we propose an adaptive selection strategy based on semantic differences between frames to improve the reliability of keyframe generation. Experimental results on the KITTI dataset indicate that, compared to other systems, the accuracy of the pose estimation has significantly improved.