To address the non-rigid deformation challenges faced by existing unsupervised deep image stitching methods in large-parallax scenes, an unsupervised image stitching approach guided by flow-field confidence and constrained by anisotropic geometric structure was proposed. The goal was to mitigate the conflict between alignment accuracy and structural integrity in deep image stitching and to enable image stitching in large-parallax scenes. First, a Flow Confidence Estimation Module (FCEM) was designed to assign confidence weights to the dense correspondence flow field generated by the context-aware layer. This module guided the network to prioritize reliable displacement signals in texture-rich regions while actively suppressing mismatches in low-confidence areas, thereby enhancing the reliability of mesh deformation. Second, an Anisotropic Geometric Structure Constraint (AGSC) was proposed, which constructed adaptive weights using Sobel edge responses and decoupled the traditional isotropic smoothness constraint into a mesh-smoothing term and a vertical-curvature term. By applying curvature penalties in the vertical direction to preserve straight-line structures while allowing horizontal deformation flexibility, the method achieves a balance between alignment accuracy and structural constraints. Experimental results on the UDIS-D dataset demonstrated that, while maintaining alignment performance comparable to that of mainstream deep stitching methods, the proposed method reduced the Line-fitting Root Mean Square Error (LRMSE) by 28.1% and the collinearity error (Eerr) by 33.4% across the full test set, effectively alleviating vertical structural bending in stitched images. Additional experiments on a 147-pair multi-scene image stitching dataset showed reductions in LRMSE and Eerr of 5.8% and 27.0%, respectively, compared to the baseline model. Both objective metrics and visual stitching results confirm the superiority of the proposed method in preserving vertical structures, verifying its geometric structure-preserving capability and cross-scene generalization performance.

With the advancement of spatial intelligence applications such as mixed reality and autonomous driving, demand for high-precision 3D reconstruction technologies in large-scale scenes has continued to increase. In such scenarios, traditional Structure from Motion (SfM) methods are computationally inefficient and prone to severe error accumulation, while existing divide-and-conquer reconstruction strategies fail to mitigate the accuracy loss caused by cumulative errors. To address these issues, a covisibility-guided hierarchical partitioning SfM method for large-scale scenes was proposed. At the global level, keyframes were selected using a greedy strategy, and GPS and visual information were fused via an adaptive covisibility weighting mechanism to rapidly generate a globally consistent sparse scene skeleton. At the local level, the scene was adaptively partitioned into several visually high-cohesion image communities based on covisibility relationships for local reconstruction. In the model fusion stage, a fusion strategy based on bidirectional consistency checking was designed to eliminate erroneous relative transformations, and all local models were aligned under global-skeleton constraints through joint optimization of multi-source transformations. Experimental results indicated that the proposed method significantly improved the efficiency and robustness of large-scale scene reconstruction and effectively suppressed cumulative errors.

After Neural Radiance Fields (NeRF), which predominantly use coordinate-based models to map spatial coordinates to pixel values, 3D Gaussian Splatting (3DGS) has emerged as an important breakthrough technique in the realm of three-dimensional reconstruction and novel view synthesis. By transforming multi-view images into millions of learnable 3D Gaussians to model an explicit scene representation, which introduces unprecedented levels of editability, and using differentiable rendering, 3DGS achieves near real-time view rendering. However, 3DGS is prone to overfitting the training views when a small number of images are available due to Gaussian splats’ local nature. Moreover, 3DGS optimizes independent splats only under multi-view color supervision, without global geometric structure. This problem becomes more pronounced as the number of images used for 3D scene optimization decreases, because sufficient images that can offer global geometric cues are unavailable. As a result, reconstructed scenes under sparse-view input often suffer from problems such as holes and floating artifacts. To address this issue, an algorithm of few-shot 3D Gaussian splatting based on monocular depth ambiguity-aware estimation was proposed to improve the quality of scene reconstruction and the effect of novel view synthesis, thereby further expanding the 3DGS application scenario. For sparse-input settings, Conditional Implicit Maximum Likelihood Estimation (CIMLE) was used to learn the multimodal distribution of the depth estimation to avoid the unimodal distribution that may be caused by the conditional GAN mode collapse. This monocular depth ambiguity-aware technique was used to extract a multi-modal depth dense point cloud for initialization of Gaussian primitives, and to introduce global geometric cues for scene reconstruction. Then, a space-carving loss was designed, along with color-based photometric loss optimization to jointly supervise and optimize the scene structure, thereby resolving the inherent uncertainty and ambiguity retained by the spatial positions of Gaussian primitives in the initialization phase. The multi-modal depth distribution was used to constrain the optimization of Gaussian primitives, with the goal of finding a globally consistent subset of patterns captured from the monocular depth ambiguity-aware distribution of each view. In this way, information from different views was fused together, since the inherent ambiguity can only be resolved through multi-view information. This allowed a common shape consistent across all views and the object surface to be “snapped” together. This approach alleviated scene overfitting under sparse views and effectively improved the quality of scene reconstruction. The experimental results showed that the average PSNR performance of the proposed three-dimensional Gaussian reconstruction algorithm was 11%-54% higher than that of the same period algorithm. At the same time, training time was greatly reduced compared with the NeRF-based algorithm, and improved scene reconstruction and novel-view-synthesis quality was achieved.

Addressing the problem that existing static 3D Gaussian Splatting (3DGS) techniques lack physical dynamic realism and the reconstructed digital assets cannot perform dynamic interactions in a physically plausible manner, a Gaussian dynamic reconstruction method based on semantic perception and the hybrid Material Point Method (MPM), termed PhysGaussian-M2, was proposed. The proposed method encompassed three core ideas: First, the Segment Anything Model (SAM) and Large Multimodal Models (LMMs) were jointly utilized to automatically infer physical constitutive parameters from visual features, thereby enabling automated parameter calibration. Second, 3D Gaussian kernels were simultaneously employed as differentiable rendering primitives and MPM Lagrangian particles, establishing a unified physical-visual representation framework and avoiding data-format mismatches in traditional pipelines. Third, within the hybrid MPM framework, independent constitutive models were assigned to different semantic regions, enabling complex physical interactions among multi-phase materials such as solids and fluids. Experimental results demonstrated that the semantic-driven parameter-estimation module accurately captured the physical laws governing different materials, with the logarithmic error between inferred parameters and standard values controlled within one order of magnitude. Multi-phase simulation experiments successfully reproduced physical phenomena such as rigid-body-fluid coupling, elastic collisions, and granular collapse, and the generated scenes exhibited high visual fidelity and physical plausibility. The proposed framework provides an end-to-end technical pathway for creating realistic and interactive physical digital-twin scenes.

Ultrasound, as a radiation-free and portable imaging modality, extends traditional 2D ultrasound utility through 3D reconstruction by integrating 2D cross-sectional slices to provide richer spatial contextual information. Existing sensor-free 3D freehand reconstruction either relies on cumbersome external localization or on deep-learning surrogates that overlook long-range spatial dependencies and poorly-textured regions. Therefore, PD²BNet, a physically-augmented, depth-synergized framework was proposed, which organically achieved complementary fusion of “internal physical interpretability and external deep generalization.” At the input stage, a three-branch parallel feature-decoupled encoder was designed to independently model grayscale texture, Canny geometric edges, and optical-flow motion vectors, thereby circumventing semantic conflicts among heterogeneous physical priors in low-level convolutions. During spatiotemporal fusion, a CNN-ConvLSTM architecture was embedded to recursively associate local topology with temporal evolution, suppressing instantaneous pose jumps. A composite physics-regularized loss was constructed, jointly incorporating velocity smoothing and physical prior regularization to significantly mitigate inter-frame jitter and accumulated drift, achieving robustness in low-texture ultrasound pullback scenarios. By combining explicit physical constraints with deep representations, accurate, sensor-free 6-DoF pose estimation was achieved, overcoming the interpretability and generalization limits of either paradigm alone. On Freehand_US_data and TUS-REC-Challenge, PD²BNet reduced final drift rate to 16.97 % and 18.24 %, respectively, with mean angular errors of merely 3.01° and 3.27°, outperforming state-of-the-art methods and achieving millimeter-level localization accuracy under typical clinical scanning protocols such as forearm examination. Quantitative and qualitative analyses confirmed that the synergy between physics augmentation and deep modeling drove the performance gains, establishing PD²BNet as a high-precision, robust algorithmic framework for clinical sensorless freehand 3D ultrasound reconstruction.

Efficiently establishing a random particle-model dataset with realistic geometric and shape characteristics is a critical issue for applications such as 3D particle recognition and particle simulation. To address the problems of insufficient morphological realism and low generation efficiency in traditional random modeling methods, a 3D random particle modeling method based on KL (Karhunen-Loeve) expansion integrated with frequency perturbation was proposed. First, three-dimensional reconstruction and spectral analysis were performed on real mineral particle surfaces to extract surface spectral features and construct particle-surface perturbation functions. Then, KL expansion was introduced to perform feature decomposition and correlation analysis on discretized particle-surface sampling data. By retaining the principal feature modes, the random generation of the primary structural morphology was achieved. Furthermore, the frequency perturbation information from real particles was incorporated into the primary structure, generating 3D random particle models with both primary structural characteristics and local perturbation features, thereby effectively improving the realism and diversity of random particle modeling. Finally, to verify the effectiveness of the proposed method, comparisons were conducted with two improved spherical harmonic analysis methods. The evaluation was performed from the perspectives of geometric characteristic parameters (length, width, and height), shape characteristic parameters (aspect ratio, sphericity, and normal perturbation angle), as well as comprehensive metrics including Wasserstein distance and model-generation efficiency. The results demonstrated that the generated particle models exhibited geometric and shape characteristic parameters that were closest to those of real particles. Moreover, the Wasserstein distance between the generated models and real particles was only 0.397 3, which was lower than 0.569 4 for the improved spherical harmonic analysis method based on fractal characteristics and 1.066 7 for the improved spherical harmonic analysis method based on genetic variation. In addition, the proposed method required less time (15.3 s) to generate the same number of particle models, achieving a favorable balance between accuracy and efficiency. The proposed method provided a new and efficient approach for generating 3D random particle models with realistic particle characteristics for applications such as 3D particle recognition and simulation.

To address the limitations of 3D Gaussian Splatting (3DGS)—namely, unnatural object-level representations, the inadequacy of explicit semantics for representing complex real-world objects, and the difficulty of ensuring cross-view consistency of pseudo-instance masks— Query-Conditioned Object Response Fields (QC-ORF) was proposed as a weakly prompted framework for constructing QC-ORF in 3D Gaussians. In 3DGS, a rendered pixel is usually produced by the continuous alpha compositing of multiple Gaussians along the viewing direction, and Gaussian primitives are not naturally aligned with real object boundaries. Therefore, directly assigning discrete object labels to individual Gaussians makes the resulting representation susceptible to interference from supporting surfaces, contact regions, occlusions, and pseudo-mask noise. Instead of relying on explicit categorical semantics or treating pseudo-masks as stable cross-view instance-level ground truth, QC-ORF models objects as query-triggered continuous Gaussian memberships. This formulation builds a continuous object response field and provides a unified object-level intermediate representation within the 3DGS framework. Specifically, the original 3DGS representation was augmented with continuous feature and foreground branches, enabling each Gaussian primitive to carry object-relevant feature information in addition to geometry, opacity, and appearance parameters. A multi-channel rendering mechanism was adopted to render RGB colors, semantic features, foreground probabilities, and depth-related information within a unified Gaussian field. During training, dense teacher features extracted from visual foundation models were distilled into the Gaussian feature branch, while pseudo-instance masks generated by SAM2 were utilized as weak prompts rather than strict object labels. Foreground-union supervision was introduced to distinguish object-related regions from the background, and single-frame intra-instance compactness was enforced to reduce local feature variance within each pseudo-instance region. Furthermore, a depth-consistency gating strategy was employed to modulate the reliability of depth-related supervision, ensuring that inconsistent monocular depth or unreliable rendered depth did not dominate the learning of the object response field. These constraints jointly facilitated the learning of stable object-relevant responses under weak and noisy supervision. During inference, object queries were instantiated by seed prompts. Positive and negative prototypes were constructed from the seed view or multiple seed views, and query responses were formulated through positive-sample attraction, negative-sample suppression, and foreground bias. The resulting target-correlated continuous object responses can be rendered as two-dimensional response maps and further decoded into downstream object-level outputs, including single-view segmentation, multi-view aggregated segmentation, and hierarchical grouping. Compared with direct hard-threshold segmentation, the continuous response field preserves the response gradients between target and non-target regions, which is beneficial for handling incomplete masks, uncertain boundaries, and supporting-surface interference. Experiments on public datasets and additionally captured real outdoor 3DGS scenes demonstrated the effectiveness of QC-ORF. The proposed method achieves favorable performance on primary object response metrics, including AUROC, AUPRC, and response gap, while exhibiting robust detail recovery in downstream segmentation tasks. In challenging scenarios characterized by incomplete training masks, complex local structures, partial occlusions, and prominent interference from supporting surfaces, QC-ORF consistently recovered cleaner and more complete target details. Visual results on real outdoor scenes further showed that stable object responses can be obtained under complex illumination, natural backgrounds, and reconstruction noise. These results indicate that stable and continuous object fields can be learned in 3D Gaussian splatting (3DGS) without explicit semantics, using only pseudo-masks as weak prompts, and that the learned response field can serve as a unified intermediate representation for object-level segmentation, hierarchical decoding, and potential scene editing.

Speech-driven 3D facial animation generation is an important research topic in computer vision and computer graphics, with broad demand in applications such as virtual digital humans, real-time interactive systems, and immersive media. The goal of this task is to generate 3D facial motions that are highly consistent with the content, rhythm, and articulation characteristics of the input speech signal. However, existing methods often rely on complex generation pipelines or multi-step sampling mechanisms, resulting in high inference latency and making them difficult to deploy in real-time streaming scenarios. To address these limitations, a real-time-oriented approach for speech-driven 3D facial animation was proposed, in which facial motion generation was formulated as an autoregressive sequence modeling problem based on discrete representations. Specifically, continuous facial motion parameters were first mapped into a compact discrete token space using VQ-VAE, thereby reducing the difficulty of sequence modeling. On this basis, a language-model-like autoregressive Transformer was employed to predict discrete facial motion tokens frame by frame under speech-conditioned constraints, enabling low-latency and streamable facial animation generation. For facial representation, the FLAME2023 parametric model was adopted in a unified manner, and audio-visual data were re-fitted using the VHAP tracker to obtain high-quality and temporally consistent facial motion parameter sequences. Experimental results demonstrated that this approach achieved competitive performance in lip synchronization, facial-expression dynamics, and head-motion consistency, while significantly outperforming existing generation methods in terms of inference efficiency and real-time performance, validating its practical applicability in real-time interactive scenarios. This research provides a feasible and scalable technical pathway for building efficient, real-time interactive 3D virtual human systems, and lays the groundwork for further incorporation of higher-level semantic controls such as speaking style and emotional states within a discrete autoregressive framework.

Gaze-based interaction, as a natural interaction modality driven by eye movements, has demonstrated great potential for improving interaction efficiency and immersive experience in target-selection tasks within virtual reality (VR) environments, due to its fast, intuitive, and hands-free characteristics. Existing studies mainly rely on input techniques such as fixation, smooth pursuit, and gaze gestures to trigger targets through predefined input parameters and interface layouts. However, fixed dwell-time thresholds and activation areas were commonly adopted during target selection, which fail to account for individual differences and dynamic user states. This limitation often leads to increased false activations and reduced triggering efficiency. To address this issue, user interaction intention in VR was investigated from an intention-driven perspective. Five representative gaze-based selection tasks were designed in a VR environment, and task-performance data, eye-movement behaviors, and subjective intention labels were systematically collected. A multi-source eye-tracking feature space was constructed based on the acquired data. On this basis, five non-sequential models based on static features (SVM, LR, RF, LightGBM, and MLP) and three sequential deep-learning models (Bi-LSTM, 1D CNN, and CNN-LSTM) were developed and evaluated. Model performance was compared using subject-independent five-fold cross-validation and an independent test set. In addition, deployment-related factors such as online inference latency and model size were jointly considered to select the optimal model. The results indicated that the LightGBM-based intention-recognition model achieved a superior balance among recognition accuracy, inference efficiency, and deployment feasibility. This study provided an effective methodological foundation for real-time intention recognition, adaptive triggering mechanisms, and personalized interaction design in gaze-controlled virtual reality interfaces.

In response to the requirements for dynamic walking ability assessment, training difficulty adjustment, and safety process monitoring in lower-limb rehabilitation for stroke patients, an embodied immersive virtual reality rehabilitation training and assessment system was constructed. The feasibility of measuring dynamic stability, body representation, and cognitive load was verified in a proof-of-concept study with healthy subjects. The system used an omnidirectional treadmill as the walking platform and employed full-body inertial motion capture to drive a first-person full-body virtual avatar. By combining head-mounted display position and orientation, event logs, and eye-tracking, the system enables the synchronous collection of data on walking kinematics, visual attention, and task behavior. A task library was built around straight walking, turning, obstacle avoidance, and dual-task walking. An index system was designed based on the root-mean-square of centroid acceleration, root-mean-square of acceleration jerk, proprioceptive localization error, pupil change, fixation/saccade characteristics, and dual-task cost. A baseline normalization threshold and hysteresis determination mechanism were introduced to form a prototype-level closed-loop of “state collection-rule determination - task adjustment-event recording”. In the lower-limb visibility experiment (N=27), the transparency of the virtual avatar’s lower limbs had a significant impact on proprioceptive localization error and embodied experience scores. The complete-invisibility condition produced higher localization errors and lower embodiment scores, while there was no significant difference in motion sickness scores. The semi-transparent condition showed a good balance between retaining body contour cues and reducing environmental occlusion. In the dual-task walking experiment (N=29), as the cognitive load increased from low to high, pupil statistics, pupil change, the number of fixations, and number of saccades generally increased, and the average fixation duration shortened, indicating that eye-movement features can reflect the load levels in dynamic walking. High gait difficulty also led to an increase in the root-mean-square of centroid acceleration and the convergence of lateral trunk swing, verifying the effectiveness of gait-difficulty manipulation. This system can integrate motor control, visual body cues, and cognitive load evidence in an immersive dynamic walking scenario, providing a methodological basis for task design, parameter adjustment, and pre-clinical evaluation of stroke rehabilitation training systems. The relevant thresholds and design strategies still need to be further verified in stroke patients.

The construction of transition surface is one of the hotspots and difficulties in CAD surface modeling technology. With the continuous improvement of surface-continuity requirements in practical engineering applications, the traditional rolling-ball method can no longer meet the actual needs of high-order smoothness in several typical scenes. Blending surfaces constructed by the existing blending methods are usually represented in the form of spline fitting, making it difficult to obtain the accurate position and derivative information, and the spline representation is often accompanied by large data storage overhead. Therefore, a G² continuity (curvature continuity) interpolation method for blending surfaces was proposed, which can accurately calculate point position information on the surface. The method focused on the analysis of the boundary constraint expression and processing of the blending surface. By constructing appropriate boundary conditions and combining them with the corresponding interpolation strategy, the G² continuity of the blending surface at the boundary was ensured, and the expression form of the blending surface was derived. This method was based on the interpolation of curves on the surface, which not only enabled simple and efficient blending-surface modeling, but also adjusted the boundary curves reasonably to meet the needs of different engineering scenarios. Finally, an accurate difference method for solving the derivatives at each point on the surface and an example of using this method to generate G²-continuous blending surface were provided. At the same time, the continuity of the obtained surface was verified by calculating the normal vector, normal curvature and generating zebra stripes. The results showed that this method can effectively meet the requirements of curvature continuity, thereby confirming its rationality.

Constrained Delaunay Triangulation (CDT) has significant practical applications in geographic information systems, 3D modeling, and engineering simulation. However, when large-scale datasets are involved, its construction process still faces challenges such as insufficient parallel efficiency, high topological update overhead, and strong platform dependency. To address these issues, a parallel constrained Delaunay triangulation algorithm based on compute shaders was proposed. Centered on the compute shader execution model, the proposed approach leveraged the fine-grained thread parallelism and high-throughput memory architecture of GPUs to achieve full-process parallelization of the CDT mesh construction workflow. In terms of algorithm implementation, the initial Delaunay triangulation was first generated through point set discretization and parallel Voronoi diagram construction. Subsequently, to meet the intersection-detection and local-reconstruction requirements during the insertion of constrained edges, a dynamic data-layout strategy based on a two-stage prefix sum was proposed to achieve compact storage and efficient indexing of intersection information. Based on this, two core data structures were designed: a dynamic intersecting-edge storage table and a dynamic constraint influence-domain index table. These were used to rapidly locate candidate edges intersecting with constraint edges and the affected triangular regions, respectively, thereby transforming the traditional global search process into local index access and significantly reducing memory access overhead. To address thread conflicts and consistency-maintenance issues in parallel topology updates, an adaptive parallel edge-flip mechanism based on local consistency was proposed. Using constraint edges as the driving unit, this mechanism confined the edge-flip process to independent thread domains. Concurrency-safe updates were achieved through atomic operations and lightweight locks, while a double-buffered queue was employed to facilitate the dynamic propagation and convergence of invalid edges, effectively avoiding concurrency issues such as redundant flips, write conflicts, and out-of-order adjacencies. Experimental results demonstrated that, for datasets containing millions of points and largescale constrained-edge sets, the proposed algorithm achieved significant improvements in computational efficiency compared to traditional serial methods and existing GPU parallel implementations, while exhibiting excellent stability and adaptability on real-world geospatial data. The research findings validated the application potential of the parallel CDT network construction method based on compute shaders in cross-platform geometric computation and high-performance graphics processing.

To address the issue that upstream design data formats in the manufacturing process of pipeline products are inconsistent, making it difficult to efficiently extract key geometric features for precise processing and assembly, a method was proposed that converted pipeline models in different formats into point-cloud data and then uniformly extracted key geometric features. First, design models in various formats were converted into point clouds to unify the data with physically scanned models. Second, a method based on boundary extraction and point-cloud segmentation was designed to obtain the boundary-circle features in the model, and the center point coordinates and radius parameters were optimized. Finally, the threshold parameters in the subsequent point-cloud processing algorithm were adaptively adjusted according to the size of the boundary circle. The pipeline model skeleton was extracted through region growing and local median iteration, and the optimization was completed in combination with cylindrical segmentation to output the axis feature parameters. Experimental results showed that compared with common skeleton-generation and geometric-segmentation algorithms, the proposed feature-extraction scheme had a wider range of applications and relatively higher accuracy. It could extract complete key geometric features from both pipeline design models with multiple branches and low-quality point-cloud models, meeting the precision requirements for calculating tooling-control commands in aviation pipeline pre-welding assembly.

The evolution of automotive electrical/electronic architectures increases the design complexity of wiring harness systems and creates bottlenecks in validation efficiency. Traditional graph traversal algorithms also face the challenges of high computational overhead and the lack of real-time feedback in large-scale harness routing and frequent local modifications. To address these challenges, an efficient computing framework based on Reserved Graph Grammar (RGG) was proposed. This framework utilized the two-layer node structure of RGG to establish a wiring harness topology model and proposed a “global topology collapse” strategy. By defining parallel reduction rules, the physical graph was efficiently transformed into a logical connection graph, reducing the algorithmic complexity from O(N²) to O(N). Experimental results showed that in the static routing scenario, this algorithm achieved a speedup ratio of 30.89 times in a 32-thread environment, with a parallel efficiency of 96.5%. Compared with the Contraction Hierarchies (CH) static acceleration algorithm, RGG reduced memory usage by approximately 50% and supported incremental updating. For dynamic design changes, a context-sensitive incremental update mechanism was introduced. In the scenarios of attribute update and topology modification, the performance advantage of incremental update became more significant as the wiring harness scale increased: when the number of nodes increased from 100 to 5 000, the speedup ratio of attribute update increased from 543 times to 4.1×10⁴ times, and the speedup ratio of topology update increased from 3.8 times to 2.2×10² times. Compared with the Ramalingam dynamic shortest path algorithm, RGG provided more stable update time and supported topology modification. Additionally, correctness verification showed that the results of this algorithm were completely consistent with those of the traditional Dijkstra algorithm (MAE=0, R²=1.0), ensuring the complete retention of physical attributes. A low-level engine with both high fidelity and real-time responsiveness was thereby provided for the analysis of ultra-large-scale wiring harnesses.

Path planning is a crucial and fundamental research direction in the field of mobile robotics. In practical engineering applications, mobile robots are often required to perform traversal tasks involving multiple target points, where the core challenge of obtaining the optimal traversal sequence lies in efficiently calculating the path cost between each pair of multi-goal points, constructing a reliable cost adjacency matrix, and further formulating the traversal problem as a Traveling Salesman Problem (TSP) for optimal solution. As a classic sampling-based motion planning algorithm, the Probabilistic Road Map (PRM) exhibits a distinct advantage of multiple query reuse after a single roadmap construction, making it suitable for batch computation of the adjacency matrix among multi-goal points. However, the conventional PRM approach suffers from an inherent trade-off between planning accuracy and computational efficiency, which severely restricts its practical application in complex multi-goal traversal scenarios. Specifically, a PRM with a high sampling density can achieve precise estimation of pairwise path costs, but it inevitably leads to excessive computational overhead and prolonged time consumption. In contrast, a PRM with a low sampling density significantly improves computational efficiency, yet it fails to ensure the feasibility of planned paths and the reliability of path cost estimation. To address the aforementioned contradiction between accuracy and efficiency, a Hierarchical Probabilistic Road Maps (HPRM) method was proposed to compute the path costs between multi-goal point pairs at differentiated precision levels. HPRM was composed of two core stages: the initial calculation stage and the recalculation stage. In the initial calculation stage, a low-sampling-rate PRM was constructed to preliminarily plan paths for all multi-goal point pairs, and an initial cost adjacency matrix was established based on the preliminarily calculated path costs. In the recalculation stage, the key multi-goal point pairs that were sensitive to the solution of the traversal order were identified and selected. For these sensitive point pairs, high-precision path replanning was performed on a high-sampling-rate PRM, and the initial adjacency matrix was updated synchronously with the refined path-cost values. Subsequently, the optimized adjacency matrix was utilized to construct the TSP model, and the optimal traversal sequence of multi-goal points was solved by TSP solvers. Finally, the global traversal path was generated by sequentially connecting each multi-goal point with the corresponding optimized local paths. To realize the rational and adaptive selection of multi-goal point pairs for recalculation, an adaptive point pair selection strategy was further proposed, integrating two sub-strategies: the path cost-based screening strategy and the multi-goal point clustering-based screening strategy. Taking the optimal clustering number of multi-goal points as the core decision criterion, this strategy employed a threshold-based judgment mechanism to identify the spatial distribution characteristics of the multi-goal point set, and adaptively selected the most appropriate sub-strategy to screen the sensitive point pairs. Comparative experiments were conducted in multiple complex simulation environments to verify the effectiveness and superiority of the proposed adaptive selection strategy. Experimental results demonstrated that, compared with the existing mainstream PRM-based multi-goal traversal methods, the path cost generated by the proposed algorithm was reduced by 0.5%~1.0%, while the computational time consumption was decreased by 20%~40%. These results indicated that the proposed HPRM algorithm effectively balanced the path planning accuracy and computational overhead, and possessed higher comprehensive computational efficiency and practical application potential in robot multi-goal path planning.

Sketch-based 3D shape retrieval has gradually become an important research direction in the field of 3D shape retrieval, mainly because it provides an intuitive mode of human-computer interaction for users. However, in real application scenarios, the number of 3D shape categories is very large and continues to grow, which makes it hard for training data to cover all potential categories, and thus traditional supervised retrieval methods based on category labels often struggle to meet the demand for new-category retrieval in open environments. This limitation motivates the study of zero-shot sketch-based 3D shape retrieval, which aims to improve the adaptability of retrieval systems to open scenarios. In this task, there are two main difficulties: addressing the modality gap between sketches and 3D shapes, and achieving effective knowledge transfer from seen categories to unseen categories in the absence of samples from unseen classes. To address these problems, a hierarchical alignment method was designed, which performs progressive alignment at both the input level and the feature level to jointly reduce cross-modal differences and support knowledge transfer. At the input level, an adversarial learning-based pseudo-view generation method was introduced to convert sketches into pseudo multi-view representations that are closer to 3D shape distributions, thereby reducing modality discrepancy from the source. At the feature level, a shared semantic embedding space was constructed, and a multi-granularity metric learning strategy was further introduced, in which semantic consistency constraints, semantic enhancement mechanisms, and class-level proxy constraints were jointly used to improve feature representation. Experimental results on the SHREC 2013 and SHREC 2014 datasets showed that the proposed hierarchical alignment framework achieved competitive performance across multiple evaluation metrics, while ablation studies further verified the effectiveness of each component, indicating that input-level alignment and feature-level semantic alignment can work together to enhance cross-modal representation learning and improve zero-shot retrieval performance.

Roller forming is a typical continuous forming process for large-scale production of aviation structural parts. Its operation involves multiple key parameters with strong coupling, such as roller speed, forming temperature and roller pressure. The fault evolution is hidden and the early signs are not obvious. Therefore, a fault prediction method based on multi-parameter threshold fusion is proposed and integrated into the digital twin system of roll forming equipment. A three-module intelligent assessment system consisting of single parameter anomaly score, multi-parameter coupling effect score and score fusion decision is constructed, and a hierarchical threshold early warning mechanism is established to realize accurate perception of equipment running status and quantification of fault risk. Among them, the single parameter anomaly scoring module divided the three-level interval for the core parameter, and quantified the parameter deviation degree through the piecewise function. The multi-parameter coupling effect module introduces coupling rules with physical significance, such as pressure/speed ratio monitoring, low-temperature/high pressure, and low-speed/high pressure, to identify composite anomalies from the perspective of a single parameter. The score fusion decision module linearly fuses each abnormal score with adjustable weight to generate a comprehensive failure probability, and combines with the dual condition trigger decision logic to realize the hierarchical early warning of equipment. Relying on the hierarchical architecture of the digital twin system, the model realizes real-time parameter acquisition, online calculation of the whole process and result visualization, and forms a closed-loop operation and maintenance process. The experimental results show that the proposed method has high prediction accuracy and good engineering adaptability in the multi-condition test set, and the performance is significantly better than the traditional single parameter threshold method. It can effectively reduce false positives and false negatives, and provide reliable technical support for predictive maintenance of roll forming equipment.

To address the challenges of complex system modeling, including the complexity of model construction and the high learning cost of SysML2.0 syntax, an intelligent SysML2.0 model generation method based on a syntax-semantic dual-layer prompting strategy was proposed. At the semantic layer, a structured semantic prompt template was constructed to standardize the representation of system compositions, behaviors, and relationships. At the syntactic layer, three prompting strategies, namely zero-shot prompting, instruction prompting, and few-shot prompting, were designed to guide Large Language Models (LLMs) in generating composition diagram and activity diagram model codes conforming to SysML2.0 syntax specifications. Meanwhile, model complexity and generation accuracy metrics were established to systematically evaluate the generation performance under different prompting strategies, model complexities, and LLM conditions. Experimental results showed that the few-shot prompting strategy achieved nearly 100% generation accuracy under different complexity levels and across different LLMs, significantly outperforming zero-shot and instruction prompting strategies. Although the generation performance of all methods decreased as model complexity increased, the few-shot prompting strategy demonstrated better adaptability and stability. The results indicated that the few-shot-driven syntax-semantic hierarchical prompting mechanism effectively alleviated semantic hallucinations and structural errors in SysML2.0 model generation by LLMs, thereby improving model generation quality without requiring large-scale annotated datasets. A feasible pathway was thus provided for the transformation of Model-Based Systems Engineering (MBSE) from manual modeling to intelligent modeling.

External well-slot extensions for jacket platforms have been widely adopted because they enable the full utilization of existing platforms to develop additional reserves in mature oil and gas fields at relatively low cost. However, most existing studies on external well-slot extensions focus on shallow-water conditions with water depths of approximately 30 m or less, whereas in ultra-deepwater environments the load-transfer mechanism of jacket-integrated systems with external well-slot extensions becomes considerably more complex. To address the limited understanding of the seismic performance mechanism of external well-slot extensions on in-service ultra-deepwater jacket platforms, a design scheme for an external well-slot extension on an in-service ultra-deepwater jacket platform was selected as the case study. A refined finite element model was established in the conventional offshore engineering analysis software SACS (Structural Analysis and Design System), and dynamic response simulations were performed based on a 200-year return-period acceleration response spectrum; additionally, discretized horizontal input directions were considered to evaluate directional effects. Connector stiffness was parameterized by modifying the elastic modulus of the connector material, and structural responses under stiffness multipliers of 0.5, 0.75, 1.0, 1.5, and 2.0 relative to the baseline were analyzed. Stress contour plots and the first five mode shapes were generated using the graphical functions of SACS to visually identify stress-concentration regions and differences in dynamic characteristics. Furthermore, focusing on sensitive members in the connection region, a mixed-effects model was introduced to quantify the relationships between connector stiffness, the maximum Unity Check (UC) value, and the associated stress components, thereby reducing the random influence arising from member-size and location variability. The results indicated that member stress levels in the connection region were significantly higher than those in surrounding areas, and that increasing connector stiffness led to an overall upward trend in the maximum UC of sensitive members, with some members exhibiting UC values exceeding 1.0 at twice the baseline stiffness. The mixed-effects analysis revealed an extremely significant positive correlation between connector stiffness and the maximum UC, and showed that the increase in maximum UC was primarily driven by the growth of shear stress and bending normal stress with increasing stiffness, whereas the effect on axial stress was not significant. Considering the direction-discretization results, seismic inputs perpendicular to the connection direction were more likely to trigger unfavorable structural responses due to displacement incompatibility induced by differences in dynamic characteristics between the jacket and the well-slot extension; excessive connector stiffness may further intensify stress concentration and reduce safety margins. It should be noted that the simulations were based on a single in-service platform design scheme and a linear analysis framework; therefore, the findings mainly reflected mechanisms and trends. Further validation across different regional seismic inputs, platform scales, and structural configurations is required, and nonlinear analyses and detailed modeling are recommended to improve engineering applicability.

Tunnel 3D reconstruction technology can be widely applied in structural defect detection, operation and maintenance state visualization, structural safety assessment, and digital twin model construction. To address the shortcomings of existing monocular vision-based tunnel 3D reconstruction technologies, such as low automation, insufficient modeling accuracy, and poor efficiency, a tunnel 3D reconstruction technique integrating drones and computer vision was proposed. On the hardware end, a tunnel inspection drone was developed, featuring a body structure made of carbon fiber and aluminum alloy, equipped with optical flow sensors, LiDAR, a three-axis gyroscope, a three-axis accelerometer, an electronic compass, and a digital barometer. With a multi-sensor fusion strategy, centimeter-level positioning and navigation were achieved in GNSS-restricted tunnel environments. The drone was also equipped with a high-performance electro-optical pod to enable real-time high-definition video acquisition and transmission. On the algorithm end, a tunnel 3D reconstruction algorithmic framework was designed, comprising video segmentation and frame extraction, camera pose and sparse point-cloud estimation using SfM, depth-map generation via the Gipuma algorithm, surface-model reconstruction based on TGV optimization, and tunnel-texture model construction using texture reconstruction techniques. Experimental results demonstrated that the tunnel inspection drone could stably operate in complex tunnel environments, capturing high-quality video data. Using video frames as the data source, the reconstructed highway tunnel 3D model achieved a root mean square reprojection error of 1.52 pixels, a maximum reprojection error of 25.44 pixels, and a total of 668 088 mesh faces. Further analysis of the impact of video-frame count and resolution on 3D reconstruction revealed that using 99 video frames as the input data achieved a balance between modeling accuracy and efficiency. When computational resources were abundant, it was recommended to upsample the original video frames before performing SfM and then use the original-sized depth maps for surface-model reconstruction. Additionally, a comparison between the proposed modeling framework and traditional solutions showed that the proposed methods exhibit higher computational efficiency and precision, as well as lower dependency on hardware resources. Furthermore, generalization tests in metro tunnels verify its engineering scalability. The proposed equipment and algorithms provide robust support for automated, accurate, and efficient tunnel 3D reconstruction.

As underground space engineering expands in scale and complexity, conventional static and discrete model delivery methods based on BIM, point clouds, or meshes suffer from slow update cycles, excessive data volumes, and visual distortions. Consequently, they are inadequate for the growing demands for dynamic visualization and precise control throughout the construction process. To address this industry challenge, a digital twin model visualization delivery framework integrating 4D Gaussian Splatting (4DGS) with Augmented Reality (AR) technology was proposed, aiming to achieve high-fidelity dynamic reconstruction and immersive delivery of underground construction models. This approach established a comprehensive technical framework encompassing temporal data acquisition, 4DGS dynamic modeling, and AR-based interactive visualization. First, by leveraging multi-view temporal imagery and 3D Gaussian Splatting (3DGS) technology, temporal attributes were integrated into the modeling of key construction nodes to construct a 4D Gaussian splatting representation of the on-site construction progress. Second, the 3DGS models were subjected to lightweight compression for deployment on mobile AR terminals. This supported virtual-real registration and overlay, temporal playback, multi-axial sectioning, and interactive real-time analysis of the digital twin across various construction phases. Experimental validation across multiple typical underground engineering scenarios demonstrated that the compressed 3DGS models achieved substantial reductions in storage footprint compared to traditional point cloud and mesh models, while significantly improving visual fidelity metrics, including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS). Furthermore, this method enabled high-fidelity, dynamic, and clear reconstruction of complex construction processes on AR devices, effectively mitigating the visual distortion issues inherent in traditional models. By integrating 4DGS and AR, the conventional static and discrete model delivery paradigm was transformed into a dynamic, continuous, and interactive immersive experience, establishing a closed-loop system encompassing perception, visualization, interaction, and feedback. This provided a robust digital foundation for construction technical briefing, process control, and full-lifecycle operation and maintenance management of underground space engineering projects.

To address practical challenges in Intangible Cultural Heritage (ICH) transmission, such as high cognitive costs, fragmented knowledge systems, and the easy loss of cultural information, a knowledge graph-driven interactive map design method was proposed using Shaanxi dough figurines as a representative case. First, a domain ontology was constructed that covered thirteen types of entity relationships, including inheritors, artworks, techniques, ritual occasions, visual symbols, social meanings, and geographical information, and spatial attribute specifications were incorporated to support geographic retrieval and spatial analysis. Second, a hybrid information extraction framework combining Aho-Corasick dictionary-matching rule templates with a domain-adaptive BERT model was proposed. This framework enabled joint extraction of entities and relations with confidence fusion on small-sample corpora, achieving F1 scores of 0.88 for entity recognition and 0.84 for relation extraction, outperforming both single-rule and single deep-learning models. Subsequently, a knowledge graph of Shaanxi dough figurines integrated with spatial indexes was established, utilizing Neo4j for triplet storage, entity linking, and associative querying. Finally, an interactive navigation map system featuring deep integration of the knowledge graph with geographic information was developed. This system enabled multi-dimensional querying and spatial analysis, including the inheritor distribution, artwork provenance, and technique transmission, while supporting dynamic maintenance of the graph data through addition, deletion, and modification operations. This study constructed an end-to-end solution encompassing knowledge extraction, modeling, storage, and visualization for ICH, effectively enhancing the precision of knowledge organization and the experience of spatial dissemination, thereby providing a reusable practical paradigm for similar ICH projects.

To address the issues of unclear cognitive mechanism and significant impact of individual psychological traits on interaction efficiency in the human-computer interaction with emotional healing robot, an emotional interaction design research was carried out for users with high and low attachment avoidance. Based on attachment theory, cognitive attention theory and human-computer interaction theory, the cognitive model of “visual attention-emotional integration” of emotional healing robots was constructed. Through a mixed experiment of 2 (anthropomorphic: high / low) ×2 (body proportion: semi-body / close-up) ×2 (interaction style: technical / social orientation), the effects of robot anthropomorphic levels, body proportion, and interaction style on emotional connection, task goal, and satisfaction of high and low attachment avoidance users were measured. Based on the eye movement experiment, the visual cognitive mechanism of high and low attachment avoidance users in early threat assessment, attention resource maintenance, and cognitive processing depth was analyzed. The SSA-BP-SHAP method was introduced to construct an emotional healing effect prediction model based on eye movement data for high and low attachment avoidance users. The study found that high and low attachment avoidance users had two different cognitive models in human-computer emotional interaction. High attachment avoidance users showed “evaluative compensation”, gaze delay when facing emotional load, and high dependence on non-core social cues such as hands to relieve stress; low attachment avoiders were characterized by “integrated immersion”, which enabled them to focus on the face to obtain emotional resonance ; there were significant differences in the influence of robot design elements on the visual attention distribution and subjective healing experience of high and low attachment avoidance users, and the prediction model could accurately quantify the relationship between visual features and healing efficacy. The research results provided a theoretical basis for the personalized design and interaction optimization of emotional healing robots for different attachment avoidance users.

In intelligent connected vehicle driving scenarios, the increasing complexity of in-vehicle instrument cluster information has led to heightened visual interference and greater cognitive workload for drivers. Based on the theory of selective visual attention, this study investigated the effects of information weighting, spatial layout, and color attributes on drivers’ attentional allocation and information acquisition efficiency, and developed corresponding design strategies aimed at improving interface readability and cognitive efficiency. This research was organized around three dimensions: “information weight-layout positioning-color characteristics.” First, the Analytic Hierarchy Process (AHP) was employed to construct a hierarchical model of in-vehicle instrument cluster information, allowing the extraction and quantification of the weight relationships among key information elements. Second, eye-tracking experiments were conducted with center-aligned and dumbbell-shaped layouts as the primary experimental conditions to analyze differences in cognitive efficiency for high-weight information, including vehicle speed, battery level and remaining range, power output, and gear status, under different spatial configurations. Third, a k-means clustering algorithm was applied to extract visual characteristics such as primary color, background color, font color, and auxiliary color, thereby summarizing the color design patterns of in-vehicle instrument cluster interfaces. Finally, based on the research findings, interface design strategies were proposed, and a comprehensive evaluation was conducted through system usability assessment, eye-tracking analysis, and user preference testing to validate the design solutions and identify the optimal scheme. The results indicated that essential driving information had the highest overall weight within the information system. Among these, vehicle speed, battery level and remaining range, power output, and gear status were identified as the most critical information elements receiving the greatest driver attention during driving tasks. When high-weight information was arranged in the visual center and its surrounding regions, both information search efficiency and recognition efficiency were significantly improved. In terms of color design, blue and blue-green color schemes are the mainstream visual styles for current instrument cluster interfaces, effectively balancing readability and visual comfort. The evaluation results further demonstrated that the proposed interface solutions exhibited good system usability. In particular, the center-aligned layout combined with a blue-green color scheme achieved superior performance in both cognitive efficiency and user preference. In conclusion, the information hierarchy, layout optimization, and color design strategies constructed based on the theory of selective visual attention effectively optimized the information structure and visual presentation of in-vehicle instrument clusters. These strategies significantly enhanced information recognition efficiency and interface readability, providing both theoretical foundations and practical guidance for the design and development of intelligent connected vehicle instrument cluster interfaces.