Search Results (2,207)

The closed-loop testing of autonomous driving systems critically depends on large-scale libraries of diverse and realistic 3D vehicle assets, yet current pipelines still rely on labor-intensive modeling or multi-view capture, making efficient construction a key bottleneck. To overcome this bottleneck and enable convenient, cost-effective 3D asset generation, we propose a semantic prior-guided framework for accurate and robust vehicle point cloud reconstruction from casually captured single-view photographs. Our framework is built on a diffusion backbone but is fundamentally driven by two forms of prior knowledge: First, geometric and appearance priors from camera-aware image features, masks, and distance-transform maps are projected onto the evolving point cloud, compensating for the severe information loss in single-view inputs. Second, we introduce distillation-style regulators—pretrained neural networks that encode vehicle type and model semantics; they act as teacher networks that impose high-level constraints on the generated point clouds, transferring rich semantic knowledge and effectively regularizing the learning process. With these priors, our model infers vehicle-specific semantics from limited observations and reconstructs high-quality 3D point cloud assets. On the 3DRealCar++ dataset, our method clearly surpasses state-of-the-art point cloud baselines in both F-score and Chamfer Distance. Full article

In satellite-denied environments, conventional localization methods struggle with rapid pose initialization due to the absence of global positioning data. To address this challenge, this study presents a high-precision pose initialization framework based on a Siamese Neural Network (SNN) and multi-stage probabilistic registration localization. First, the SNN improved by Convolutional Block Attention Module (CBAM) matches features between real-time radar point clouds and prior map slices, producing candidate positions based on similarity scores. Then, Adaptive Monte Carlo Localization (AMCL) performs probabilistic matching among these candidates to identify the correct slice and refine the position accuracy from tens of meters to meter-level, along with an approximate orientation estimate. Finally, the Normal Distributions Transform (NDT) is applied for point cloud registration, achieving centimeter-level pose estimation. The proposed method is evaluated on self-collected medium-scale and large-scale maps. Experimental results show that the SNN effectively identifies the correct map slice, which is further refined by AMCL and NDT to achieve centimeter-level position accuracy and sub-degree orientation accuracy. The multi-stage method achieves localization success rates of 99% on both 200 × 100 m and 300 × 200 m regions, with distance RMSEs of 0.175 m and 0.348 m, and orientation RMSEs of 0.149° and 0.437°, respectively. Evaluations on the KITTI dataset further demonstrate robust initialization performance in complex outdoor environments. The proposed framework provides a reference for high-precision pose initialization in large-scale satellite-denied scenarios. Full article

Aiming at the problems of extremely scarce measured samples and significant domain shift between simulated and measured data in automatic target recognition (ATR) of air targets for spaceborne radar, this paper proposes an inverse synthetic aperture radar (ISAR) image recognition method for air targets combining physics-driven data augmentation guided by detection prior information with domain adversarial transfer learning. First, the mapping relationship between scattering point projection and ISAR images is established by using the target 3D point cloud and radar observation geometric priors, and a 2D sinc kernel function is introduced for energy distribution rendering. Then, under the unsupervised transfer learning paradigm, aiming at the distribution inconsistency between augmented data (source domain) and unlabeled simulated data (target domain), this paper designs a cross-domain recognition task experiment including six types of typical aircraft targets, and compares the cross-domain recognition performance of three transfer learning methods (model fine-tuning, deep domain confusion (DDC) and domain-adversarial neural networks (DANN)) on the target domain. Meanwhile, t-distributed stochastic neighbor embedding (t-SNE) visualization is used to analyze the feature distribution alignment ability of the models. Simulation experiments show that the DANN model with a dynamic inversion coefficient introduced in the gradient reversal layer (GRL) achieves a recognition accuracy of 99.5% on the unlabeled target domain, which is significantly superior to the model fine-tuning and DDC methods. Moreover, it makes the feature distributions of source and target domain samples highly overlapping, and maintains a strong inter-class discriminability while eliminating the domain shift. The proposed scheme provides a physically interpretable and robust technical path for few-shot radar target image recognition. Full article

In industrial quality control, the precise identification of severe structural weld defects is paramount. Traditional 2D image-based detection methods are susceptible to illumination and texture interference, while high-precision 3D laser scanning solutions are costly and impractical for large-scale deployment. To achieve reliable geometric defect detection at low cost, this paper proposes a detection framework based on monocular depth estimation and a dual-attention point cloud network. First, YOLOv8 is employed for rapid region of interest extraction, and an advanced monocular depth estimation model generates 3D pseudo-point clouds containing geometric information. Secondly, addressing the challenge of distinct spatial orientation features in missed weld defects that are prone to confusion, this paper introduces a dual-attention-enhanced point cloud classification network named DA-PointNet++. This model embeds dual-attention modules within the PointNet++ backbone network, enhancing key feature representation in both the channel and spatial dimensions. Experimental results demonstrate that this approach achieves an accuracy of 93.67% and a recall rate of 90.51% in a unified binary classification task for general weld defect detection, effectively identifying both normal welds and complex missed weld defects. Compared to PointConv, Dynamic Graph Convolutional Neural Network (DGCNN), and mainstream Point Cloud Transformer, this method significantly reduces false negative rates while maintaining low computational costs, offering a cost-effective solution for industrial automation. Full article

Accurate and efficient 3D reconstruction of trees is of paramount importance for studying forest spatial structures and dynamic resource patterns, optimizing forest management, protecting environments, and analyzing carbon cycles. Currently, Light Detection and Ranging (LiDAR) remains the dominant method for generating 3D models of forest scenes. However, with advancements in computer vision, photogrammetry has emerged as a crucial tool for forest inventory and 3D reconstruction due to its cost-effectiveness. Nevertheless, in practical forestry applications, traditional photogrammetry often suffers from low reconstruction efficiency and poor quality during feature extraction and matching. These issues stem from the complex structure of forest scenes, severe occlusion, and repetitive texture patterns. To address these challenges, this paper proposes an improved 3D tree reconstruction approach based on images, integrating deep learning-based methods. In the sparse reconstruction stage, we utilize the ALIKED (A LIghter Keypoint and descriptor Extraction network with Deformable transformation) algorithm and construct an image pyramid to extract multi-scale robust features. Furthermore, by combining the LightGlue matching algorithm with a neighborhood search constraint strategy, we enhance the stability of camera pose recovery while reducing redundant computations. Experimental results demonstrate that our method outperforms traditional algorithms in both accuracy and robustness regarding image matching. Compared to baseline models, the proposed approach increases the number of feature points by approximately 50% with a more widespread distribution, improves matching accuracy by 4% to 8%, and achieves a 100% image registration rate. Consequently, under the condition of maintaining equivalent re-projection errors, the subsequent sparse point clouds exhibit an average track length increase of 0.6 to 1.4 and a density increase of up to 1.2 times. Notably, this method effectively mitigates artifacts and spurious reconstructions caused by pose drift in forest photogrammetry. Full article

Accurate road segmentation is a core perceptual technology for autonomous driving, but faces two challenges: (1) ambiguous road boundaries caused by insufficient modeling of contextual information relationships in CNN-based networks and (2) inadequate LiDAR-camera fusion due to modality gaps between heterogeneous sensors. To mitigate these limitations, this paper proposes a novel approach, named LiDAR-Camera High-Resolution Feature Fusion Network (LC-HR2FNet), a multi-cross-stage fusion model designed for road segmentation. Firstly, a new type of pseudo-LiDAR-Image representation is generated via an early-level fusion strategy and data complementation. Sparse point clouds are transformed into dense LiDAR-Image data and then concatenated with RGB channel maps to form complementary multi-modal data inputs. Subsequently, a modified HRNet backbone integrated with cross-stage feature fusion is constructed to strengthen information interaction across different branches and enhance the modeling of contextual relationships. Additionally, a dilated feature collection model is designed to collect multi-scale confidence scores for pixel-wise class determination. Experiments on the KITTI road benchmark demonstrate that the proposed method achieves a MaxF of 97.39% on UMM_ROAD and an average of 96.28% across all urban scenarios, demonstrating superior performance and robustness. Full article

To address the issue where conventional initial registration methods combined with the Iterative Closest Point (ICP) algorithm are highly sensitive to initial parameters such as point cloud density and pose—often resulting in degraded accuracy or even misregistration—this paper proposes a Lidar point cloud registration method integrating triangular mesh features. First, Crust Triangulation is employed to construct a triangular mesh from the input point cloud. Then, the keypoint extraction based on triangular meshes, termed KE-TM, is introduced. Subsequently, a Triangle Feature Histogram (TFH) is constructed as the feature descriptor. Based on this, an initial alignment method grounded in triangular meshes, referred to as TM-IA, is developed to achieve coarse registration of point clouds. Finally, the ICP algorithm is applied to refine the alignment. Comparative experiments conducted on incomplete Bunny point clouds demonstrate that the proposed KE-TM maintains a higher keypoint repeatability under reduced point cloud density. The combined TM-IA and ICP registration method can achieve rotation errors within 1° and translation errors within 1 mm under small initial pose deviations, while also maintaining robust performance under larger initial misalignments. Compared with traditional methods, the proposed method significantly reduces sensitivity to initial parameters and improves the accuracy. This method has certain practical significance for the precise alignment of 3D point clouds. Full article

To improve the data quality of underwater point clouds acquired by UAV-borne LiDAR bathymetry, a scanline-based sliding window filtering method is proposed based on an analysis of scanline data characteristics. Scanline data of underwater point clouds are first extracted from raw point clouds, and the radius outlier removal algorithm is employed to eliminate outliers. Taking the acquisition time of scanline points as the X-axis and elevation as the Y-axis, a 3D problem is simplified into a 2D representation, and a sliding window is constructed along the scanline. Robust least-squares fitting is applied within the window. The median absolute deviation of the fitting residuals is adopted to calculate the terrain feature values for quantifying the terrain complexity, followed by an adaptive filtering threshold determination according to terrain feature values. Fine filtering of the individual scanlines is performed using a point-by-point sliding window. Experimental results demonstrate that the proposed method is adaptable to various terrain conditions, achieving a noise recall rate ≥ 96%, an overall filtering accuracy ≥99%, and an F1-score ≥ 0.9. Particularly, the precision rate in flat-water areas reached 97.37%. Overall, the proposed filtering method effectively separates noise points while preserving detailed terrain features and supports UAV-borne LiDAR bathymetry for mapping complex shallow-water regions. Full article

Point cloud scene flow estimation aims to capture the three-dimensional motion of each point in a sequence of point clouds. Although progress has occurred in this field, existing methods often face significant challenges. In particular, two key issues persist: the absence of corresponding local information from the source point cloud to the target, preventing correct feature matching, and the presence of highly similar adjacent structures in target regions, which leads to ambiguous correspondences due to indistinguishable point features. To address these problems, this paper introduces a novel self-supervised method for point cloud scene flow estimation. Theoretically, we establish a new framework that integrates discriminative feature learning with probabilistic flow refinement. A new network architecture, SetConv++, is designed to learn more discriminative point feature representations, enhancing differentiation in similar structures. Additionally, a refinement module uses the random walk algorithm to adjust initial flow estimates. This approach reconstructs low-confidence flows with high-confidence surrounding ones, reducing missing correspondence issues. Crucially, a new flow smoothing loss term ensures local consistency while suppressing error propagation—a fundamental limitation in existing methods. Through comprehensive experiments on the KITTI Scene Flow dataset, our method demonstrates superior performance. It significantly outperforms existing self-supervised approaches across multiple standard evaluation metrics. Specifically, on the KITTI Scene Flow dataset, our method reduces the Endpoint Error (EPE) by 13.6% (from 0.0411 to 0.0355) and improves Accuracy Strict (AS) by 2.43 percentage points (from 92.68% to 95.11%) compared to baseline self-supervised approaches, while also reducing the outlier rate (Out) by 1.5 percentage points. This advancement not only provides a robust theoretical framework for handling ambiguous correspondences but also enables more reliable and efficient downstream applications—such as autonomous driving perception systems requiring real-time motion accuracy in complex scenes. Full article

LiDAR point clouds provide accurate and direct representations of spatial locations and geometric structures of objects in 3D space, making them essential for applications such as target recognition in autonomous driving and 3D reconstruction in smart cities. However, large-scale point clouds pose challenges, including massive data volume, uneven density distribution, and complex object structures. Existing point-based and graph-based semantic segmentation networks often suffer from limitations such as loss of local contextual information, over-reliance on local graph construction, and insufficient modeling of relationships between neighboring points. To address these issues, we propose PG-Net, a novel network that integrates discrete point distribution features with local graph structural information. The framework includes: (1) a point branch equipped with a Local Adaptive Feature Augmentation (LAFA) module to extract efficient local features; (2) a graph branch featuring a Dynamic Graph Feature Aggregation (DGFA) module, which explicitly models relationships among points in local graphs and adaptively balances a point’s intrinsic features with its neighborhood context; and (3) fuses local features from both branches, allowing their complementary strengths to enhance feature representation, a process further promoted by a New Aggregation Loss Function. Experiments on the Toronto3D and S3DIS datasets show that PG-Net achieves overall accuracy (OA) of 97.69% and 89.87%, and mean Intersection-over-Union (mIoU) of 83.51% and 73.22%, respectively. Comparative and ablation studies against advanced methods such as RandLA-Net, BAAF-Net, and LACV-Net demonstrate the effectiveness and robustness of our approach. By jointly exploiting discrete point distribution and local graph structural relationships, PG-Net effectively leverages the complementary strengths of its dual-branch design, offering a reliable solution for efficient and accurate large-scale point cloud semantic segmentation. Full article

Point cloud registration in environments lacking rich textures or containing repetitive structures remains highly susceptible to misalignments. The core challenge lies in balancing the demand for extracting highly distinctive local features with the computational cost of global context modeling. In this paper, we propose a robust registration framework that efficiently combines rotation-equivariant geometric representations with state space models of linear complexity to mitigate feature ambiguity and mismatch. First, a multivariate geometric encoding mechanism is embedded within convolutional layers, enhancing local feature distinctiveness under strict rotation equivariance by explicitly leveraging surface properties. Second, to efficiently establish long-range spatial dependencies, we replace standard dense attention with a hybrid geometry-state aggregation module. This module integrates local geometric self-attention with the Mamba architecture, strengthening focus on overlapping regions without the quadratic computational burden. Finally, we optimize the generated correspondences through a physically consistent hypothesis generator to compute reliable rigid transformation results. On standard benchmarks, our framework demonstrates exceptional robustness to ambiguous matches, achieving a 96.3% registration recall on the 3DMatch dataset and outstanding accuracy on the KITTI dataset. Full article

High-rise LiDAR scanning produces vertically sparse point clouds where upper-layer defects are hardest to detect due to inverse-square ranging law (1/r²) density gradients, noise contamination, and complex geometries. This paper presents PC-TowerNet, a physics-aware AI pipeline that achieves state-of-the-art reconstruction through sequential modules: (1) 50D geometric feature classification outperforming CloudCompare SOR (100% accuracy vs. 91.3% retention); (2) Physics-Constrained VAE (PC-VAE) recovering 28.7 ± 2.1% upper density vs. 8.3 ± 1.7% standard VAE; (3) multi-modal PointNet++/GNN/Transformer fusion; and (4) Bayesian uncertainty maps (ECE = 0.042 ± 0.008). Synthetic tower evaluation (10 × 5 seeds) demonstrates 48.9% surface smoothness improvement and 38.2% volume error reduction over tuned RANSAC baselines, with clear paths to real-data validation. Full article

In the field of subgrade compaction quality assessment, accurate volume measurement of excavated pits is hindered by non-uniform point cloud distribution, environmental noise interference, and complex irregular boundary features. To address these challenges, this paper proposes a robust volume detection framework that integrates adaptive point cloud refinement and morphological discrimination. First, a pose normalization method employing RANSAC plane fitting and rigid body transformation corrects the spatial orientation of the raw point clouds. To balance data redundancy removal with feature preservation, a gradient adaptive simplification strategy based on local density feedback and K-nearest neighbor estimation is developed. Subsequently, a cross-sectional area calculation model utilizing piecewise-cubic polynomial fitting is proposed to mitigate boundary noise and accurately reconstruct irregular contours. Furthermore, a dynamic outlier removal mechanism based on the Median Absolute Deviation (MAD) and sliding windows is introduced to eliminate non-physical geometric fluctuations. Finally, the total volume is aggregated using a hybrid strategy of Simpson’s rule and a frustum compensation operator. Experimental results on simulated pits with typical topological defects demonstrate that the proposed algorithm outperforms traditional methods, achieving an average relative volume error of less than 0.8%. This approach significantly improves the robustness and precision of sensor-based automated subgrade compaction quality measurement. Full article

In close-range monocular visual measurement and cooperative target pose estimation, conventional planar targets are constrained by viewpoint changes and are prone to perspective distortion. Although spherical targets provide omnidirectional observability, their PnP-based pose estimation may still suffer from large errors under limited fields of view and sparse feature observations. To address this issue, this paper proposes an integrated visual measurement framework covering both high-precision spherical target construction and robust pose estimation. First, a composite marker layout based on adaptively scaled latitude–longitude topology is designed. To suppress cumulative distortion caused by long-sequence multi-view rigid registration, a center-to-pole point-cloud stitching strategy is developed, and multiple observations are fused using geometric-consistency weighting to accurately reconstruct the feature-point coordinate system of the target. Second, a joint optimization method is proposed by combining feature-point reprojection error with a contour center consistency constraint. Specifically, the theoretical contour center is predicted from the analytical projection model of the sphere and constrained to agree with the observed contour center fitted from the image. In addition, an SQPnP-based sequential reinitialization mechanism is introduced to improve robustness under sparse-point observations. Simulation results demonstrate that the proposed method achieves higher accuracy and robustness under continuous pose changes, sparse feature points, and different noise levels, compared with EPnP, EPnP+LM, LM, and SQPnP, while real-image experiments further demonstrate its practical feasibility. Full article

Aiming at the inherent bottlenecks of traditional smoke detection technologies in high and large-volume building scenarios, this paper conducts research on an early fire smoke detection method for high and large-volume spaces based on Light Detection and Ranging (LiDAR). A special experimental platform was independently designed to obtain the physical characteristics of smoke particles from standard smoldering fires. Combined with the optical scattering and reflection interaction mechanism between laser and particulate matter, the theoretical feasibility of LiDAR for smoke detection was systematically verified. Smoke irradiation experiments were carried out in the full detection distance, and the LiDAR point cloud characterization characteristics of smoldering smoke were clarified. A special smoke detection algorithm based on point cloud features was designed, a LiDAR smoke detection system was built, and multi-condition comparative experiments with traditional photoelectric smoke detection methods were carried out in a full-scale laboratory. The experimental results show that the LiDAR-based smoke detection method proposed in this paper has significant advantages over traditional detection methods in terms of alarm response speed, detection coverage, and height adaptability. This research provides a brand-new technical path and reference for the theoretical research and engineering application of early fire warning technology for high and large-volume buildings. Full article