3d reconstruction and mapping

To address the insufficient accuracy of LiDAR odometry for wheeled mobile robots in planar environments (e.g., warehouses, hospitals), this paper proposes a robust multi-sensor odometry method integrating kinematic constraints. The core contribution is the first explicit incorporation of the platform’s kinematic model into a point-to-point ICP optimization framework, coupled with a dynamic weighting mechanism that adaptively increases the contribution of wheel encoders during LiDAR feature degradation. The method combines an improved ICP algorithm, real-time nonlinear optimization, and weighted fusion of heterogeneous sensor data. Experimental evaluation demonstrates that the proposed approach achieves significantly higher localization accuracy than conventional LiDAR-only and wheel-only odometry methods in large-scale warehouse and outdoor scenarios. It has been successfully deployed in Dexory’s global warehouse robot navigation system.

This paper addresses the challenge of accurately localizing the center of checkerboard targets in unstructured, sparse, and high-noise 3D point clouds—particularly those acquired by low-cost LiDAR sensors. To this end, we propose a robust measurement framework specifically designed for such data. Our method integrates 3D template matching, synthetic-data-driven target modeling, robust preprocessing of real-world point clouds, and sub-pixel-level center estimation into an end-to-end pipeline. A key innovation is the first application of color-enabled Iterative Closest Point (ICP) to template matching on unstructured point clouds, thereby overcoming the strong reliance of conventional approaches on point cloud ordering and low noise levels. Experimental results demonstrate sub-pixel localization accuracy on synthetic data and validate the method’s feasibility and robustness in practical applications—including point cloud registration, long-term structural monitoring, and multi-sensor fusion—using real LiDAR acquisitions.

To address the challenges of limited computational resources and inaccessibility of raw sensor data on VR/AR devices—which severely degrade the real-time performance of conventional SLAM systems—this paper proposes a sparse SLAM framework based on geometric projection of 3D static meshes. Unlike traditional approaches, it introduces mesh-based geometric projections as robust, calibration-free visual features, eliminating dependence on raw sensor inputs and enabling lightweight pose estimation. The method comprises geometric feature extraction, sparse keyframe optimization, and mesh-guided pose solving, tightly integrated into the VR runtime. Evaluated on real-world VR hardware, the system achieves over 30 FPS, reduces computational overhead by 67%, and lowers localization error by 41% compared to ORB-SLAM2. These results significantly enhance the feasibility and real-time performance of SLAM deployment under resource-constrained conditions.

To address the degradation of ICP-based LiDAR odometry accuracy caused by dynamic objects, non-overlapping observations, and sensor noise, this paper proposes a robust ICP variant evaluation and enhancement framework. Methodologically: (1) a dynamic point cloud filtering module explicitly detects and removes moving objects; (2) a self-vehicle blind-zone compensation mechanism improves registration reliability in low-observability regions; (3) a 2D projection-based objective-function visualization method quantitatively reveals how each interference factor affects ICP convergence behavior and pose estimation error. Experiments on KITTI and SynLIO demonstrate that the proposed framework significantly enhances both localization accuracy—reducing average translational error by 23.6%—and runtime stability across diverse ICP variants in complex urban scenes. The framework provides an interpretable evaluation tool and practical enhancement strategies for robust LiDAR odometry.

Iterative Closest Point (ICP) registration fails in geometrically degenerate environments—e.g., long corridors or large open spaces—where LiDAR point clouds lack sufficient geometric constraints. Method: This paper proposes and empirically evaluates a degradation-aware registration framework. It conducts the first large-scale comparative study of active versus passive degradation-mitigation strategies; introduces truncated singular value decomposition (TSVD), inequality-constrained optimization, and linear/nonlinear Tikhonov regularization; and develops a sensitivity analysis model tailored to the ICP least-squares step. Experiments span real-world robotic field expeditions and high-fidelity simulations. Contribution/Results: When reliable external pose priors are unavailable, active degradation mitigation proves indispensable. The proposed soft-constraint approach—tuned heuristically—demonstrates significantly improved robustness and accuracy over standard ICP in ill-conditioned scenarios, establishing a new benchmark for degenerate-environment LiDAR registration.

This work addresses the high communication overhead and low data efficiency in multi-robot collaborative SLAM, which often stems from reliance on low-level feature matching. The authors propose a distributed SLAM framework based on scene graph matching that leverages RGB-LiDAR fused point clouds for semantic segmentation, extracting discrete objects and their boundaries to construct scene graphs. Notably, inter-robot loop closures are achieved solely by exchanging object labels and centroids, eliminating dependence on raw feature descriptors. Integrated with a multi-stage communication scheme and distributed optimization, the method significantly reduces communication load while preserving localization and mapping accuracy, as demonstrated in both simulated and real-world experiments with legged robots across indoor and outdoor environments.

Conventional geometric accuracy monitoring of unstructured point cloud data (PCD) suffers from error accumulation, high computational cost, and artifact generation due to reliance on registration and mesh reconstruction. Method: This paper proposes a preprocessing-free, intrinsic geometry-aware monitoring method that directly extracts intrinsic geometric features from raw PCD via Laplacian eigenmaps and geodesic distance computation. It incorporates an unsupervised threshold selection mechanism and a dual-strategy feature learning framework to enable robust defect identification. Contribution/Results: Experimental results demonstrate that the method achieves efficient and accurate detection of diverse geometric defects—including dents, bulges, and misalignments—without registration or reconstruction. It maintains high precision (>96% F1-score) and strong stability across heterogeneous manufacturing scenarios (e.g., additive manufacturing, CNC machining), reducing monitoring time by up to 78% compared to state-of-the-art approaches while eliminating reconstruction-induced artifacts. The framework significantly enhances both efficiency and reliability in industrial PCD-based quality inspection.

Large-scale mobile laser scanning (MLS) point cloud registration in urban street scenes faces challenges including non-uniform point density, noise contamination, and severe occlusion. To address these, this paper proposes an adaptive segmentation and optimization-based registration framework. First, a novel semi-spherical segmentation (SSC) preprocessing method is introduced, leveraging orthogonal planar features to achieve optimal trajectory segmentation. Second, a voxelized plane selection strategy is incorporated to enhance robustness and avoid local minima. Third, a plane-voxel-guided generalized ICP (PV-GICP) algorithm is developed for efficient and precise registration. Evaluated on real-world MLS data from Munich’s city center, the framework achieves a mean registration accuracy of 0.009 m while reducing computational time by over 50%. This significantly advances the automation level of large-scale urban 3D modeling and dynamic monitoring.

This work addresses the challenge of effectively leveraging underutilized or deprecated panoramic RGB and LiDAR log data to construct high-quality 3D Gaussian Splatting (3DGS) digital twins. To this end, the authors propose a reusable multimodal data processing pipeline that enables robust initialization for 3DGS. The pipeline features deterministic spatial anchoring from equirectangular projection (ERP) to cubemap representation, PRISM point cloud downsampling guided by color stratification, and a two-stage registration process combining FPFH-based global alignment with ICP-based fine refinement. This approach achieves efficient and accurate fusion of RGB and LiDAR data, significantly enhancing geometric consistency and rendering fidelity of 3DGS in complex scenes compared to vision-only baselines. The method successfully unlocks the practical value of archived sensor logs, demonstrating their potential for high-fidelity 3D reconstruction.

Traditional SLAM systems struggle to simultaneously deliver high-fidelity geometric reconstruction and critical semantic information in complex environments, limiting robotic situational awareness in applications such as disaster assessment and industrial inspection. To address this, this work proposes a pixel-level multimodal approach that fuses visible and infrared imagery, projecting real-time LiDAR point clouds onto the fused image plane. High-temperature targets are segmented using thermal channels and embedded as a temperature-aware semantic layer within the 3D map. This method achieves, for the first time, pixel-aligned integration of thermal semantic information into real-time 3D semantic mapping, significantly enhancing immediate detection and labeling of high-temperature objects. The approach demonstrates practical utility in rapid disaster response and predictive industrial maintenance scenarios.