Cabin segment assembly critically determines the performance and safety of high-end equipment in aerospace, aviation, shipbuilding, and energy sectors. This paper presents a hybrid serial-parallel dual-arm manipulator for cabin docking. Each arm provides four motion types (axial, vertical, radial, circumferential), and the modular design enables adaptation to varied cabin sizes with coordinated six-degree-of-freedom movement. The kinematic model is established using the Denavit-Hartenberg method, the workspace is analyzed via Monte-Carlo simulation, and B-spline curves ensure smooth trajectory planning. To address control-accuracy limitations, transmission-error models for joint drives and a Stribeck friction model are integrated, identifying transmission-induced harmonic disturbances and velocity-dependent friction as dominant error sources. A quasi-sliding mode controller (QSMC) is designed based on Newton-Euler dynamics to actively suppress disturbances, compensate for errors, and eliminate chattering typical of conventional sliding-mode control. Compared to classical sliding mode control (SMC), the quasi-sliding mode control (QSMC) strategy significantly improves the trajectory tracking performance. The maximum absolute tracking errors for joint displacement and the peak fluctuations in velocity tracking are effectively suppressed, with reductions ranging from 97.70% to 99.98%. Laser-tracker-based repeat positioning tests confirm that the accuracy of all joints and the end-effector surpasses design specifications (e.g., joints within +0.045 mm, end-effector within +0.047 mm). Further experimental evaluations verify that the dual-arm cabin docking system achieves industrial-grade performance, with attitude-angle repeatability errors below 0.1 degrees, positional accuracy better than 0.2 mm, and attitude-control accuracy better than 0.15 degrees-demonstrating its high precision and reliable engineering applicability.

Infrared tiny target detection is of great value in fields such as military reconnaissance and security early warning, but faces challenges including low signal-to-noise ratio (SNR), performance-efficiency tradeoffs, and detection-false alarm compromises in complex dynamic scenarios. To solve these questions, we propose a novel spatio-spectral-temporal progressive (SSTP) algorithm, integrating spatial, spectral, and temporal features for infrared tiny target detection in cluttered scenes. First, it adopts an anisotropic gradient difference detection algorithm to construct a spatial candidate target set based on the anisotropic radiation characteristics of target neighborhoods. Then, we use the isolation penalty adaptive clustering algorithm to obtain boundaries via outlier-enhanced clustering, and design a multilateral context filling algorithm to generate suspected regions and fill internal boundary information. In addition, we develop an adaptive nonlinear geometric filter for point screening using nonlinear structural features, apply a multiscale wavelet energy filter to capture high-frequency features, and utilize a target-background local difference measurement algorithm to extract regional independence for screening. Based on the proposed single-frame detection method, a multidimensional feature fusion-based dynamic target tracking algorithm is employed to extract moving targets. Experiments show that on multiframe datasets DSAT and single-frame datasets SIRST, the proposed method significantly outperforms mainstream algorithms, achieving detection rates of 98.75% and 98.23% as well as false alarm rates of 2.56x10-6 and 10.86x10-6 , respectively. The algorithm not only performs well in multiframe detection, but also has good performance in single-frame detection. It thus provides a solution with high robustness and real-time performance for infrared early-warning systems in complex environments.

The classification of aerial images can provide crucial data sources for Earth observation. Although conventional neural network classification models based on large sample training have made significant strides in accuracy, the high costs of obtaining large samples and training are constraints on the widespread application of these methods. Inspired by few-shot learning methods, this paper proposes SCAFFNet (Spatial-Channel Attention and Fourier Transform Fusion Network) to address the issue of few-shot remote sensing scene image classification. By combining attention mechanisms and Fourier transform techniques, SCAFFNet enhances feature extraction and classification performance with limited data samples. Experiments on the UC Merced and WHU-RS19 datasets validate the effectiveness of SCAFFNet, demonstrating that the proposed method improves classification accuracy compared to traditional methods and showing its potential for practical applications in remote sensing image analysis.

Multimodal Large Language Models (MLLMs) are increasingly expected to solve structured perception tasks that require visual recognition, language-to-object binding, object cardinality preservation, and precisely localized grounding and segmentation outputs. However, existing group-relative reinforcement learning methods provide only response-level supervision, creating a granularity mismatch for structured multi-object prediction: a single advantage is broadcast to all tokens in a response, without distinguishing individual box contributions. To address this mismatch, we propose MCR-GRPO, a marginal contribution assignment framework that derives box-level credit directly from each sampled response. Specifically, Marginal Contribution Reward (MCR) estimates each predicted box's contribution through a leave-one-out comparison, measuring how the matched set value changes when the box is removed from the response. After within-response normalization, records that improve the set value receive positive credit, while redundant or harmful ones are suppressed. To make marginal attribution stable and informative, we further introduce a Continuous Matched Set Value Evaluator that integrates permutation-invariant matching, count-aware normalization, and graded localization. MCR-GRPO maps normalized box-level marginal advantages to the token spans that generated each box, preserving GRPO's response-level comparison while enabling box-aware optimization of structured multi-object grounding. Experiments across REC, DOD, segmentation, and counting benchmarks show state-of-the-art performance over prior GRPO-based baselines.

Infrared small target detection (IRSTD) remains challenging due to the inherently low signal-to-noise ratio (SNR), complex background clutter, and indistinct target boundaries. To tackle these interconnected issues, we propose a novel dual-branch network based on the segment anything model (DBSAM) that couples background suppression with explicit edge structure modeling. The architecture consists of two specialized components: 1) within the encoder, our proposed adaptive wavelet-based background suppression (AWBS) modules employ a stationary wavelet transform (SWT) with adaptive thresholding to simultaneously denoise encoder features and enhance target saliency through multiscale decomposition and 2) in parallel, the edge-aware fusion branch (EAFB) processes the Sobel-derived edge maps utilizing dynamic snake-like receptive fields to precisely capture fine-grained boundary structures. Extensive experiments on four benchmark datasets demonstrate that DBSAM outperforms existing methods across multiple evaluation metrics, including detection probability (P-d), false-alarm rate (F-a), and intersection over union (IoU). Ablation studies further validate the contributions of both AWBS and EAFB to enhance overall performance.

Object detection in real-world environments is inherently challenging due to various factors, such as poor illumination conditions, which can significantly hinder detection accuracy. To overcome the above challenges, we propose a multispectral feature-fusion object detection framework by designing a dual-modal backbone based on YOLOv8. Within the backbone, we propose a Feature Reconstruction Fusion module based on the Swin Transformer, which reconstructs dual-modal features by considering the intensity of spatial information and learns complementary inter-modal information. Moreover, we design multi-scale auxiliary branches during the training stage to complement the single-modal gradient information at different levels, strengthening the model's ability to recognize multi-scale targets. Our method is tested on several public datasets, including the FLIR-aligned, LLVIP, and M3FD datasets. The results show that our network not only achieves remarkable accuracy in detecting small-sized targets but also outperforms other state-of-the-art networks in mean average precision.

In dynamic visual scenes, salient object detection faces critical challenges in maintaining temporal consistency and spatial accuracy, particularly manifested in insufficient spatiotemporal feature integration and inadequate modeling of long-term temporal dependencies. To tackle these challenges, we introduce the Multi-Scale Temporal Memory Network (MSTMNet), a novel framework that synergizes hierarchical spatiotemporal modeling with adaptive memory mechanisms. MSTMNet employs a dual-encoder architecture to separately extract features from the current query frame and historical memory frames, ensuring rich contextual representation. A dynamic memory bank with an attention-guided elimination policy selectively retains historically significant features based on relevance and freshness, while a Multi-Temporal Feature Fusion (MTFF) module aggregates multi-scale features across sequential temporal steps to capture both short-term motion cues and long-range dependencies. Subsequently, within the Spatial Feature Fusion (SFF) module, an enhanced self-attention mechanism refines low-level features by propagating semantic consistency from high-level representations. By combining memory-enhanced features with query features through a hierarchical decoder, MSTMNet achieves precise saliency maps. Comprehensive experiments on public datasets validate that MSTMNet surpasses transformer-based methods and outperforms established semi-supervised and optical-flow-dependent techniques in accuracy.

The increasing use of unmanned aerial vehicles (UAVs) for remote sensing image segmentation has revolutionized applications such as smart agriculture, disaster monitoring, and urban planning. However, current methods often rely on fully supervised learning, requiring extensive labeled data and struggling with zero-shot capabilities for unseen categories. To address these challenges, we propose AerialCLIP, a lightweight open-vocabulary method for real-time semantic segmentation of UAV-captured remote sensing images, based on the widely-used vision-language model (VLM), i.e., CLIP. While CLIP excels in zero-shot predictions, its large parameter size prevents direct application on UAV platforms with limited computational resources. Therefore, we introduce a two-stage architecture, incorporating a saliency-based mask proposal generation (SMPG) module to efficiently generate foreground class masks. Additionally, we apply knowledge distillation to reduce the computational overhead of CLIP, enabling deployment on resource-constrained edge devices. Our extensive experiments across multiple UAV-based remote sensing datasets-UAVid, UDD5, and VDD-demonstrate that AerialCLIP achieves significant improvements, with an average mIoU of 44.1%, 51.2%, and 45.9%, respectively, while reducing model parameters by over 50%, showcasing both high accuracy and parameter efficiency.

Given the fact that visual-inertial odometry (VIO) is faced with the challenges of localization drift in the long run, we utilize drift-free Ultra-Wideband (UWB) measurements to eliminate accumulated errors in VIO. Existing UWB-VIO fusion methods are mostly constrained by the accuracy of prior UWB anchor positions. However, in large-scale localization scenarios, the precise locations of UWB anchors are difficult to obtain, and the offline calibration process is complex, significantly limiting flexibility. In this paper, we firstly design a lightweight initialization method based on a dual sliding window structure, which can rapidly obtain initial guesses for the UWB anchor coordinates. After that, we further propose a joint estimation system to refine the anchor coordinates while estimating the correction for VIO. The system combines filter-based and optimization-based methods, which mainly consists of an initialization module and a nonlinear estimator module. The filter in the initialization module provides optimization initial values and covariances, and mutually, the optimization results from the nonlinear estimator provide priors for the filter. Finally, the performance of our proposed approach is verified through both public datasets and real-world experiment. Our project, along with our dataset, has been open-sourced in the form of ROS package and ROS bag.

Target localization is a key technology for unmanned aerial vehicle (UAV) applications in various fields, such as target tracking and task planning. However, the accuracy of UAV localization is significantly affected by systematic and random errors in attitude data, and the nonlinearity of the measurement model, together with the unknown distribution of measurement noise. To achieve robust and precise localization in long-distance oblique scenarios based on dynamic platforms, this article proposes a Gaussian Mixture Filter-incorporated self-attention (SA) Residual Neural Network (GMAR) algorithm for target localization. Firstly, an end-to-end SA residual neural network (SA-ResNN) model is built to accurately model both systematic and random errors in attitude angle. The SA mechanism is innovatively introduced to enhance the global feature representation capability of the residual module. Then, the Gaussian mixture (GM) filter utilizes a GM model to model the prior and posterior probability density functions, which can effectively capture the uncertainty in the state probability density function under nonlinear measurement models and enhance the robustness of the localization system. Finally, simulations and flight experiments demonstrate that the proposed GMAR algorithm can significantly improve the localization accuracy and robustness of ground targets in long-distance oblique scenarios.

Existing planners commonly rely on cell-based environment representation approaches, which oversimplify the geometry and face the challenges of low spatial resolution and high maintenance costs. Inspired by cloth simulation techniques, we propose a novel environment representation method called Cloth Simulation Field (CSF). By simulating the interactions of cloth particles under gravity in the forward direction, obstacles and free areas can be quickly delineated. Furthermore, the distribution of particles provides natural auxiliary information for generating collision-free trajectories. Therefore, we further propose a trajectory generation method compatible with CSF. Different from traditional optimization-based planners, which often rely on repulsive forces from obstacles to push trajectories away and consequently introduce non-convexity, the proposed method leverages the unique gradient in CSF to transform repulsive forces into attractive forces, enabling trajectory generation to be formulated as a convex optimization problem, which is further solved in closed form, significantly reducing computational costs. Finally, the proposed method is validated through simulations and real-world experiments. The comparative experiments demonstrate that our planner outperforms state-of-the-art planners in terms of trajectory quality and replanning time. Our project has been open-sourced as a ROS package.

3D path planning is a field that has attracted much attention in the quadrotor research. Aiming at the problem of online trajectory generation when waypoints change during flight, this paper proposes a polynomial trajectory online adjustment algorithm. By redistributing local time, new waypoints are added to the local trajectory to perform a closed-form solution. This algorithm allows the quadrotor to adjust its flight trajectory in time in response to changes of waypoints. At the same time, the spatial and temporal continuity of the trajectory is guaranteed while the dynamic constraints is satisfied. The simulation experiment results verify the real-time performance and stability of the proposed method for local trajectory replanning in unknown environments.

Person search aims at finding the target person from the panoramic image with the multiple camera views. It can be widely used in many areas such as public security and intelligent video surveillance. However, person search is a challenging task due to the cluttered background, occlusion, and person pose variations. We present a person search algorithm based on improved SeqNet to address these challenges. On the basis of SeqNet, an attention mechanism is utilized to extract global features and local features, which are fused with the features extracted by SeqNet to enhance the discrimination of person features. In addition, a feature update strategy is designed based on IoU groups to solve the problem of gradually weakening historical features during the feature update process. The comprehensive experiments are executed on CUHK-SYSU and PRW datasets, the results show that our proposed method is effective and superior to most existing works.