Semantic Scene Completion (SSC) is a task that simultaneously predicts the occupancy and semantic labels of the environment. Compared with separate processing, SSC leverages the coupled nature of scene completion and semantic segmentation. Although this multitask integration can utilize complementarity and correlation between tasks, it also increases the training difficulty. To address this, in this letter, we propose a Semantic Decoupling based Semantic Scene Completion (SD-SSC) network from a single depth image. The semantic segmentation task is decoupled from the semantic scene completion task, and we use 2D and 3D semantic supervision to simplify the scene completion task and improve SSC performance. Specifically, our network first performs 2D semantic segmentation on the depth image and transforms features into 3D voxel space as semantic priors. Then, the 3D SSC is performed based on the voxel features and the flipped Truncated Signed Distance Field (f-TSDF). We use multi-scale 3D semantic supervision to further enhance the semantic information and fuse semantic and geometric features through the Planar Attention Fusion Module (PAFM) to obtain accurate SSC results. The proposed SD-SSC network achieves state-of-the-art performance on the NYU dataset (51.1% mIoU) and the NYUCAD dataset (61.9% mIoU) among all single depth-image based methods. It is even better than most RGB-D fusion-based SSC methods.

Wide field-of-view (FoV) LiDAR sensors provide dense geometry across large environments, but existing LiDAR-inertial-visual odometry (LIVO) systems generally rely on a single camera, limiting their ability to fully exploit LiDAR-derived depth for photometric alignment and scene colorization. We present Omni-LIVO, a tightly coupled multi-camera LIVO system that leverages multi-view observations to comprehensively utilize LiDAR geometric information across extended spatial regions. Omni-LIVO introduces a Cross-View direct alignment strategy that maintains photometric consistency across non-overlapping views, and extends the Error-State Iterated Kalman Filter (ESIKF) with multi-view updates and adaptive covariance. The system is evaluated on public benchmarks and our custom dataset, showing improved accuracy and robustness over state-of-the-art LIVO, LIO, and visual-inertial SLAM baselines.

Visual saliency prediction seeks to mimic human visual attention. Despite significant advancements in this field through deep learning (DL) techniques, several challenges remain unresolved: (1) The degree and scale of saliency vary significantly across different regions and objects within an image. Traditional approaches, such as image pyramids, skip-layers, or inception modules, lack flexibility and provide insufficient perceptual diversity. (2) Transformers and non-local blocks have proven effective in encoding long-range contextual information for visual saliency prediction. However, they come with considerable computational and memory requirements. To address these issues, we propose a lightweight Global Context Attention (GCA) module for visual saliency prediction, which captures long-range contextual information by leveraging tensor decomposition theory with low computational overhead. Moreover, we propose the Perceptual Diversity (PD) module, which effectively captures multi-scale features through diverse receptive fields. By repeatedly applying the PD module, the network acquires a broad spectrum of receptive fields, enabling it to capture salient information at various scales and levels of detail. Extensive experiments on diverse benchmarks demonstrate that our approach achieves a superior trade-off between computational cost and accuracy compared to existing state-of-the-art (SOTA) methods.

Fine segmentation of sub-conductors plays a critical role in enabling precise condition assessment for high-voltage transmission corridor inspections. With the widespread applications of bundle conductors in modern power grids, Unmanned Aerial Vehicle Light Detection and Ranging (UAV LiDAR) systems have become indispensable for acquiring sub-conductor point clouds, valued for their operational efficiency and measurement accuracy. However, achieving accurate segmentation in complex power scene point clouds remains challenges due to severe noise interference, high feature similarity among targets, minimal spacing between sub-conductor, and uneven point cloud distribution. To overcome these limitations, this study proposes a novel sub-conductor segmentation algorithm with three core innovations. Firstly, an adaptive multi-scale geometric feature descriptor is designed for the linear span-direction distribution of transmission lines to capture candidate points while mitigating noise interference. Secondly, non-linear mapping strategy using spatial geometric regularities across conductor configurations to reduce dimensionality. Thirdly, an adaptive multi-density perception algorithm is proposed to resolve point cloud density heterogeneity for precise segmentation. Validated across diverse terrains, conductor types, and noise levels, the method achieves an average precision, recall, and F1-score all exceeding 95 %. These results demonstrate its significant potential to advance intelligent inspection capabilities and enhance management efficiency for high-voltage transmission corridors. (c) 2025 COSPAR. Published by Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

Line feature matching is a fundamental task in computer vision, with wide applications in 3D reconstruction, image registration, and visual positioning. Traditional methods, often reliant on local feature descriptors or pre-defined geometric models, frequently struggle in the presence of large-scale scaling, non-rigid transformations, or sparse texture scenes. This paper proposes a robust line feature matching method that combines spatial clustering of point features with point-line geometric invariants. First, an improved RFM-SCAN algorithm is employed to spatially cluster initial point feature matches, effectively purifying them by enforcing motion consistency. Subsequently, for the line feature matching, the core innovation is introduced: a dual-invariant construction comprising a point-to-line distance ratio invariant and a direction vector invariant. High-accuracy line correspondences are then achieved by leveraging these dual invariants in conjunction with rigorous geometric consistency constraints. Experimental results demonstrate that our method maintains high matching accuracy and robustness across various challenging conditions—including scale and rotation changes, occlusion, illumination variation, and low-texture scenes—without relying on GPU acceleration or large-scale training data.

Subway shield tunnels have emerged as the preferred solution for urban transportation due to their convenience and safety. Constructed using prefabricated concrete segments, these tunnels exhibit structural stability. However, the segment joints and bolt holes are prone to groundwater infiltration under prolonged external stress, potentially compromising the lifespan of the shield tunnels. Consequently, effective detection methods are imperative to ensure the safe operation of these tunnels. Accurate data acquisition and precise extraction of leakage features are critical for detecting leakages in subway tunnels. This research introduces Efficient Geometry Enhanced Point Transformer Network (GEPT-Net), an innovative point cloud semantic segmentation network designed specifically for detecting tunnel leakage. GEPT-Net leverages the observation that leakages predominantly occur at segment joints and bolt holes, characterized by distinct geometric features and lower intensity. The network incorporates Fast Point Feature Histograms (FPFH) to effectively capture these geometric features from the input data. Additionally, we introduce a point cloud serialization technique utilizing space- filling curves, enabling the network to perceive a greater number of points within the same computational power, thereby balancing efficiency and accuracy. The Geometry Enhanced Channel Attention (GECA) Block is introduced to enhance the interaction between FPFH feature channels and intensity channels, enhancing the precise localization of leakage areas. Furthermore, the Lovasz Hinge Loss is employed to address the issue of extreme class imbalance. We constructed a tunnel leakage point cloud dataset, named S3DIS_leakage, comprising approximately 1,600 m between two stations, to train and evaluate the performance of our network. Experimental results demonstrate that GEPT-Net achieves superior performance in tunnel leakage semantic segmentation, attaining approximately 85 % mean Intersection over Union and 89 % accuracy for leakage classes, outperforming cutting-edge 2D and 3D networks by at least 12 %. Moreover, GEPT-Net maintains a remarkable balance between segmentation accuracy and computational efficiency, rendering it viable for practical engineering applications. This study not only establishes a robust approach for tunnel leakage detection but also paves the way for future research on the comprehensive segmentation of shield tunnel components. The proposed framework is available from the following github repository: https://github.com/jdjiang312/GEPT-Net.

Freespace detection plays an important role in autonomous driving. In recent years, deep learning based freespace detection methods have performed well in urban scenes. However, for off-road scenes, freespace detection poses significant challenges due to the complexity of the scenes and the lack of clear edges. The existing methods have not effectively fused LiDAR data and camera images. In this paper, we propose a Pyramid Cross-Modal Feature Fusion Network (PCMF2-Net) for off-road freespace detection. The dense depth maps are concatenated with RGB images and used as input along with surface normal maps. The dual branch CNN-Transformer encoder combines convolutional neural networks and transformers to extract local and global features from RGBD images and surface normal maps, respectively. Then, in the pyramid cross-modal feature fusion module, the multi-scale and multimodal encoder features are fused in a top-down manner. In addition, we also use an edge segmentation task and a two-step training strategy to further improve performance. Experiments on the off-road freespace detection dataset (ORFD) demonstrate that the proposed PCMF2-Net achieves a competitive result of 93.9% IoU at a speed of 23 Hz.

Line feature matching is a crucial aspect of computer vision and image processing tasks, attracting significant research attention. Most line matching algorithms predominantly rely on local feature descriptors or deep learning modules, which often suffer from low robustness and poor generalization. In response, this paper presents a novel line feature matching approach grounded in point–line invariants through spatial invariant relationships. By leveraging a robust point feature matching algorithm, an initial set of point feature matches is acquired. Subsequently, the line feature supporting area is partitioned, and a constant ratio invariant is formulated based on the distances from point to line features within corresponding neighborhood domains. Additionally, a direction vector invariant is also introduced, jointly constructing a dual invariant for line matching. An initial matching matrix and line feature match pairs are derived using this dual invariant. Subsequent geometric constraints within line feature matches eliminate residual outliers. Comprehensive evaluations under diverse imaging conditions, along with comparisons to several state-of-the-art algorithms, demonstrate that our proposal achieved remarkable performance in terms of both accuracy and robustness. Our implementation code will be publicly released upon the acceptance of this paper.

Line feature matching is a fundamental and extensively studied subject in the fields of photogrammetry and computer vision. Traditional methods, which rely on handcrafted descriptors and distance-based filtering outliers, frequently encounter challenges related to robustness and a high incidence of outliers. While some approaches leverage point features to assist line feature matching by establishing the invariant geometric constraints between points and lines, this typically results in a considerable computational load. In order to overcome these limitations, we introduce a novel Bayesian posterior probability framework for line matching that incorporates three geometric constraints: the distance between line feature endpoints, midpoint distance, and angular consistency. Our approach initially characterizes inter-image geometric relationships using Fourier representation. Subsequently, we formulate the posterior probability distributions for the distance constraint and the uniform distribution based on the constraint of angular consistency. By calculating the joint probability distribution under three geometric constraints, robust line feature matches are iteratively optimized through the Expectation–Maximization (EM) algorithm. Comprehensive experiments confirm the effectiveness of our approach: (i) it outperforms state-of-the-art (including deep learning-based) algorithms in match count and accuracy across common scenarios; (ii) it exhibits superior robustness to rotation, illumination variation, and motion blur compared to descriptor-based methods; and (iii) it notably reduces computational overhead in comparison to algorithms that involve point-assisted line matching.

Visual exploration is a task in which a camera-equipped robot seeks to efficiently visit all navigable areas of an environment within the shortest possible time. Most existing visual exploration methods rely on a static camera fixed to the robot's body to control its own movements. However, coupling the orientation of camera with robot's body limits the extra degrees of freedom to obtain more visual information. In this work, we adjust the camera orientation during robot motion by using a novel camera view planning (CVP) policy to improve the exploration efficiency. Specifically, we reformulate the CVP problem as a reinforcement learning problem. However, two new challenges need to be addressed: 1) determining how to learn an effective CVP policy in complex indoor environments and 2) figuring out how to synchronize it with the robot motion. To solve the above issues, we create a reward function considering factors such as exploration area, observed semantic objects, and the motion conflicts between the camera and the robot's body. Moreover, to better coordinate the policies of the camera and the robot's body, the CVP policy takes the body actions and the egocentric 2D spatial maps with exploration, occupancy, and trajectory information into account to make motion decisions. Experimental results show that after using the proposed CVP policy, the exploration area is expanded by 21.72% and 25.6% on average in the small-scale indoor scene with few structured obstacles and large-scale indoor scene with cluttered obstacles, respectively.

The visual-inertial simultaneous localization and mapping (VI-SLAM), which integrates data from monocular or stereo cameras, has garnered significant attention and development. The RGB-D camera, capable of capturing both color and depth images simultaneously, can perceive a comprehensive view of the surroundings. To fully leverage two types of measurement information from the RGB-D camera and inertial measurement unit (IMU) sensor for accurate pose estimation, we propose a new VI-SLAM algorithm, VID-SLAM, that effectively couples the RGB-D camera with the IMU. In our proposal, we first develop an adaptive point feature detection approach that rapidly detects and tracks sufficient point features. This approach uses adaptive nonmaximum suppression and the KD-Tree algorithm to ensure a homogeneous distribution of point features. Second, we incorporate line features into the pose estimation module of the simultaneous localization and mapping (SLAM) algorithm. By screening line features based on the geometric properties of vanishing points, we ensure that the detected lines align with the edges of scene objects as early as possible. Beyond the 2-D reprojection error of line features, we introduce a new error term that leverages the geometric constraints of plane normal vectors formed by matched line features and the optical center of the RGB-D camera; furthermore, we estimate the pose of the RGB-D camera by loosely coupling point-line visual features with IMU preintegration measurements. In the back end of VID-SLAM, we tightly couple the point-line feature error model with the IMU preintegration to jointly optimize the camera pose. Extensive qualitative and quantitative comparisons demonstrate that our VID-SLAM algorithm achieves robust performance and comparable accuracy.

In the construction of large-scale water conservancy and hydropower transportation projects, the rock mass structural information is often used to evaluate and analyze various engineering geological problems such as high and steep slope stability, dam abutment stability, and natural rock landslide geological disasters. The complex shape and extremely irregular distribution of the structural planes make it challenging to identify and extract automatically. This study proposes a method for extracting structural planes from UAV images based on Geo-AINet ensemble learning. The UAV images of the slope are first used to generate a dense point cloud through a pipeline of SfM and PMVS; then, the multiple geological semantics, including color and texture from the image and local geological occurrence and surface roughness from the dense point cloud, are integrated with Geo-AINet for ensemble learning to obtain a set of semantic blocks; finally, the accurate extraction of structural planes is achieved through a multi-semantic hierarchical clustering strategy. Experimental results show that the structural planes extracted by the proposed method perform better integrity and edge adherence than that extracted by the AINet algorithm. In comparison with the results from the laser point cloud, the geological occurrence differences are less than three degrees, which proves the reliability of the results. This study widens the scope for surveying and mapping using remote sensing in engineering geological applications.

Now, most existing dynamic RGB-D SLAM methods are based on deep learning or mathematical models. Abundant training sample data is necessary for deep learning, and the selection diversity of semantic samples and camera motion modes are closely related to the robust detection of moving targets. Furthermore, the mathematical models are implemented at the feature-level of segmentation, which is likely to cause sub or over-segmentation of dynamic features. To address this problem, different from most feature-level dynamic segmentation based on mathematical models, a non-prior semantic dynamic segmentation based on a particle filter is proposed in this paper, which aims to attain the motion object segmentation. Firstly, GMS and optical flow are used to calculate an inter-frame difference image, which is considered an observation measurement of posterior estimation. Then, a motion equation of a particle filter is established using Gaussian distribution. Finally, our proposed segmentation method is integrated into the front end of visual SLAM and establishes a new dynamic SLAM, PFD-SLAM. Extensive experiments on the public TUM datasets and real dynamic scenes are conducted to verify location accuracy and practical performances of PFD-SLAM. Furthermore, we also compare experimental results with several state-of-the-art dynamic SLAM methods in terms of two evaluation indexes, RPE and ATE. Still, we provide visual comparisons between the camera estimation trajectories and ground truth. The comprehensive verification and testing experiments demonstrate that our PFD-SLAM can achieve better dynamic segmentation results and robust performances.

In the study of RGB-D SLAM (Simultaneous Localization and Mapping), two types of primary visual features, point and line features, have been widely utilized to calculate the camera pose. As an RGB-D camera can capture RGB and depth information simultaneously, most RGB-D SLAM methods only utilize the 2D information within the point and line features. To obtain a higher accuracy camera pose and utilize the 2D and 3D information within points and lines better, a novel geometric constraint model of points and lines (PL-GM) using an RGB-D camera is proposed in this paper. Our contributions are threefold. Firstly, the 3D points and lines generated by an RGB-D camera combining with 2D point and line features are utilized to establish the PL-GM, which is different from most models of point-line SLAM (PL-SLAM). Secondly, in addition to the 2D re-projection error of point and line features, the constraint errors of 3D points and lines are constructed and minimized likewise, and then a unified optimization model based on PL-GM is extended to the bundle adjustment model (BA). Finally, extensive experiments have been performed on two public benchmark RGB-D datasets and a real scenario sequence. These experimental results demonstrate that our method achieves a comparable or better performance than the state-of-the-art SLAM methods based on point and line features, and point features.

RGB-D camera is a new type of sensor, which can obtain the depth and texture information in an unknown 3 D scene simultaneously, and they have been applied in various fields widely. In fact, when implementing such kinds of applications using RGB-D camera, it is necessary to calibrate it first. To the best of our knowledge, at present, there is no existing a systemic summary related to RGB-D camera calibration methods. Therefore, a systemic review of RGB-D camera calibration is concluded as follows. Firstly, the mechanism of obtained measurement and the related principle of RGB-D camera calibration methods are presented. Subsequently, as some specific applications need to fuse depth and color information, the calibration methods of relative pose between depth camera and RGB camera are introduced in Section 2. Then the depth correction models within RGB-D cameras are summarized and compared respectively in Section 3. Thirdly, considering that the angle of the view field of RGB-D camera is smaller and limited to some specific applications, we discuss the calibration models of relative pose among multiple RGB-D cameras in Section 4. At last, the direction and trend of RGB-D camera calibration are prospected and concluded.

RGB-D SLAM (Simultaneous Localization and Mapping) generally performs smoothly in a static environment. However, in dynamic scenes, dynamic features often cause wrong data associations, which degrade accuracy and robustness. To address this problem, in this paper, a new RGB-D dynamic SLAM method, PLD-SLAM, which is based on point and line features for dynamic scenes, is proposed. First, to avoid under-over segmentation caused by deep learning, PLD-SLAM combines deep learning for semantic information segmentation with the K-Means clustering algorithm considering depth information to detect the underlying dynamic features. Next, two consistency check strategies are utilized to check and filter out the dynamic features more reasonably. Then, to obtain a better practical performance, point and line features are utilized to calculate camera pose in the dynamic SLAM, which is also different from most published dynamic SLAM algorithms based merely on point features. The optimization model with point and line features is constructed and utilized to calculate the camera pose with higher accuracy. Third, enough experiments on the public TUM RGB-D dataset and the real-world scenes are conducted to verify the location accuracy and performance of PLD-SLAM. We compare our experimental results with several state-of-the-art dynamic SLAM methods in terms of average localization errors and the visual difference between the estimation trajectories and the ground-truth trajectories. Through the comprehensive comparisons with these dynamic SLAM schemes, it can be fully demonstrated that PLD-SLAM can achieve comparable or better performances in dynamic scenes. Moreover, the feasibility of camera pose estimation based on both point features and line features has been proven by the corresponding experiments from a comparison with our proposed PLD-SLAM only based on point features.

With extensive application of RGB-D cameras in robotics, computer vision, and many other fields, accurate calibration becomes more and more critical to the sensors. However, most existing models for calibrating depth and the relative pose between a depth camera and an RGB camera are not universally applicable to many different kinds of RGB-D cameras. In this paper, by using the collinear equation and space resection of photogrammetry, we present a new model to correct the depth and calibrate the relative pose between depth and RGB cameras based on a 3D control field. We establish a rigorous relationship model between the two cameras; then, we optimize the relative parameters of two cameras by least-squares iteration. For depth correction, based on the extrinsic parameters related to object space, the reference depths are calculated by using a collinear equation. Then, we calibrate the depth measurements with consideration of the distortion of pixels in depth images. We apply Kinect-2 to verify the calibration parameters by registering depth and color images. We test the effect of depth correction based on 3D reconstruction. Compared to the registration results from a state-of-the-art calibration model, the registration results obtained with our calibration parameters improve dramatically. Likewise, the performances of 3D reconstruction demonstrate obvious improvements after depth correction.