Registration aims at merging multiple scans to cover all scenes of a large environment. Thus, it is crucial to many civil infrastructure applications based on three-dimensional (3D) models. However, in many real-world scenarios, it is necessary to align point clouds with low-density or small overlaps. It is difficult to extract stable features and enough features for registration, whether keypoint features or overall posture features, under this condition. Existing methods cannot solve this problem well. This work proposed an end-to-end registration network that can self-adaptively focus on the overlap. The network learned to directly encode posture information from the overlapping area instead of using sparse keypoint correspondences, which makes the network more generalized and efficient. This work also proposed a self-supervised overlapping detector as an extension module to expand the use of this network to align large-scale point clouds of indoor building environments. The proposed detector is compatible with any registration approaches to promote their accuracy and efficiency further. The proposed network was experimentally demonstrated to outperform the state-of-the-art methods in registering sparse and low-overlapping point clouds, with higher robustness to point density and overlap ratio change. The proposed detector can reliably detect the overlapping area and empower the network to accurately align the sparse and low-overlapping point clouds of the large-scale indoor scene, thus simplifying and promoting laser scanning practices in civil infrastructure applications.
The pedestrian stride-length estimation is a crucial piece of personal behavior data for many smartphone applications, such as health monitoring and indoor location. The performance of the present stride-length algorithms is suitable for simple gaits and single scenes, but when applied to sophisticated gaits or heterogeneous devices, their inaccuracy varies dramatically. This paper proposes an efficient learning-based stride-length estimation model using a smartphone to obtain the correct stride length. The model uses adaptive learning to extract different elements for changing and recognition tasks, including Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) modules. The direct fusion method maps the eigenvectors to the appropriate stride length after combining the features from the learning modules. We presented an online learning module to update the model to increase the SLE model’s generalization. Extensive experiments are conducted with heterogeneous devices or users, various gaits, and switched scenarios. The results confirm that the proposed method outperforms other state-of-the-art methods and achieves an average 4.26% estimation error rate in various environments.
Point cloud-based retrieval for place recognition is essential in robotic applications like autonomous driving or simultaneous localization and mapping. However, this remains challenging in complex real-world scenes. Existing methods are sensitive to noisy, low-density point clouds and require extensive storage and computation, posing limitations for hardware-limited scenarios. To overcome these challenges, we propose LWR-Net, a lightweight place recognition network for efficient and robust point cloud retrieval in noisy, low-density conditions. Our approach incorporates a fast dilated sampling and grouping module with a residual MLP structure to learn geometric features from local neighborhoods. We also introduce a lightweight attentional weighting module to enhance global feature representation. By utilizing the Generalized Mean pooling structure, we aggregated the global descriptor for point cloud retrieval. We validated LWR-Net’s efficiency and robustness on the Oxford robotcar dataset and three in-house datasets. The results demonstrate that our method efficiently and accurately retrieves matching scenes while being more robust to variations in point density and noise intensity. LWR-Net achieves state-of-the-art accuracy and robustness with a lightweight model size of 0.4M parameters. These efficiency, robustness, and lightweight advantages make our network highly suitable for robotic applications relying on point cloud-based place recognition.
18 December 2025 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Point cloud filtering serves as a core step in point cloud data processing, which is critical for constructing high-precision Digital Elevation Models (DEMs) and conducting terrain monitoring. However, existing filtering methods encounter two key challenges. First, they exhibit strong reliance on threshold parameters and limited adaptability, necessitating manual parameter adjustment to suit different scenarios—this often results in misclassification or under-classification of ground and non-ground points. Second, their adaptability to complex terrains is inadequate: in regions with drastic slope variations or dense buildings, issues such as the loss of terrain details or confusion between terrain points and feature point clouds frequently arise, making it difficult for a single method to meet the requirements of diverse application scenarios. To address this problem, this article proposes a multi-scale mesh fitting filtering based on adaptive clustering segmentation and gradient compensation. First, the existing clustering segmentation algorithm is improved to propose an adaptive clustering segmentation method. Then, by combining Delaunay triangulation and local feature parameters, an outlier cluster is constructed. On the basis of the clustering segmentation results and the distribution of outliers, the building feature areas are effectively separated. After that, grids are constructed by separating the point clouds of the building clusters. The point clouds within the grids are judged by using planes and slope angles to remove most non-ground points, and establishes a secondary grid. The elevation threshold is determined on the basis of the fitted grids and gradient compensation. Finally, carrying out a secondary check according to the cluster edges obtained by clustering. The experimental results demonstrate that the proposed method has wide applicability without the need to switch feature parameters, and its filtering accuracy is significantly higher than that of existing methods in most point cloud scenarios.
11 February 2026 in IEEE Internet of Things Journal
Self-supervised monocular depth estimation leverages photometric consistency across views as supervisory signals, yet inherently suffers from compromised geometric fidelity at object boundaries, surface smoothness, and cross-view consistency due to 3-D structural information loss during 2-D projection. This article introduces Diff3D-Net, a novel framework integrating multilevel 3-D geometric constraints to overcome these limitations. Key innovations include: a fixed-parameter RPM-Net point cloud registration module generating geometrically grounded supervision; a pose consistency constraint that directly optimizes the pose network using RPM-Net-refined transformations to suppress error accumulation; a dual-path photometric constraint adaptively fusing reconstructions from original and optimized poses to enhance low-texture robustness; and a multiview 3-D structural consistency constraint enforcing global point alignment via minimum residual selection. Evaluations on KITTI and DDAD show Diff3D-Net’s state-of-the-art performance, achieving 0.094/4.220 Abs. Rel/root mean squared error (RMSE) on KITTI. The model also exhibits strong cross-domain generalization in zero-shot settings, outperforming competitors on the Cityscapes, NYU Depth V2, and DrivingStereo datasets without fine-tuning, highlighting its Internet of Thing (IoT) application potential. Despite higher training costs, its inference achieves an optimal speed–accuracy tradeoff, with a lightweight variant reaching 51.73 FPS—sufficient to enable real-time applications such as UAV navigation on resource-limited IoT edge devices. Ablations confirm each constraint’s contribution, and qualitative results demonstrate robustness in dynamic scenes and complex geometries.
23 March 2026 in IEEE Transactions on Mobile Computing
Visible Light Positioning (VLP) has emerged as a promising indoor localization technology owing to its high accuracy, low power consumption, and lighting compatibility. The Received Signal Strength (RSS)-based multi-anchor VLP pre-serves these advantages while having drawn considerable research attention due to its simple implementation and high reliability. However, its large-scale deployment encounters challenges at every stage: inefficient anchor calibration, limited model-based ranging performance, and robustness reduction from undetected gross errors. To address these issues, we propose a VLP and inertial navigation system fusion framework comprising an anchor position estimation module, a distance estimation module, and a fusion positioning module. For anchor calibration, a LiDAR-inertial odometry-based calibration scheme enhanced by a twolayer optimization strategy is introduced, which provides prior knowledge of anchor positions for the whole system. To improve the performance of the RSS-based ranging method, a hybrid Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN)-Bidirectional Long Short-Term Memory (Bi-LSTM) network with an embedded distance quality assessor is developed, achieving over 50% higher accuracy than model-based baselines. It delivers precise distance estimates and their validity labels for measurement updates in subsequent fusion positioning. Additionally, a two-stage error detection mechanism filters low quality observations by combining network-generated usability labels with prior-state estimates. The system consistently attains decimeter-level positioning across various trajectories, meeting the needs of diverse Internet of Things applications.