Search Results (3,735)

Simultaneous localization and mapping (SLAM) is the basis for intelligent robots to explore the world. As a promising method for 3D reconstruction, 3D Gaussian splatting (3DGS) integrated with SLAM systems has shown significant potential. However, due to environmental uncertainties, errors in the tracking process with 3D Gaussians can negatively impact SLAM systems. This paper introduces a novel dense RGB-D SLAM system based on 3DGS that refines Gaussians through sub-Gaussians in the camera coordinate system. Additionally, we propose an algorithm to select keyframes closely related to the current frame, optimizing the scene map and pose of the current keyframe. This approach effectively enhances both the tracking and mapping performance. Experiments on high-quality synthetic scenes (Replica dataset) and low-quality real-world scenes (TUM-RGBD and ScanNet datasets) demonstrate that our system achieves competitive performance in tracking and mapping. Full article

To address the issues of low recognition rates and poor detection accuracy for road negative obstacles caused by insufficient feature representation, we propose a novel detection framework: the Negative Road Obstacles Segmentation Network (NROSegNet). The detection performance of the algorithm is improved through a data enhancement strategy based on feature splicing and an adaptive feature enhancement module. Specifically, the data augmentation strategy introduces negative obstacle features into other datasets through geometric transformations and random splicing, effectively increasing the diversity of training data. This can solve the problem of an uneven distribution of data features while improving the performance of the model in capturing illumination changes and local details. The framework further adopts a dynamic multi-scale feature enhancement module to improve the perception of local details and global semantics. A robust multimodal data fusion mechanism and edge-aware optimization strategy are designed to effectively alleviate the problems of noise interference and boundary blur. The experimental results show that the NROSegNet proposed in this paper achieves 70.6 and 83.0 in mIoU and mF1, respectively, which is 2.8 percentage points and 2.9 percentage points higher than other methods. The results fully demonstrate its excellent performance in precise segmentation and boundary detail processing. Full article

Traditional methods for estimating tea yield mainly rely on manual sampling surveys and empirical estimation, which are labor-intensive and time-consuming. Accurately estimating fresh tea production in different seasons has become a challenging task. It is possible to estimate the seasonal yield of tea at the field scale by using the spatial resolution of 10 m, 5-day revisit period and rich spectral information of Sentinel-2 imagery. This study integrated Sentinel-2 images and uncrewed aerial vehicle (UAV) RGB imagery to develop six regression models at the field scale, which were employed for the estimation of seasonal and annual fresh tea yields of the Yunlong Tea Cooperatives in Yixiang Town, Pu’er City, China. Firstly, we gathered fresh tea production data from 133 farmers in the cooperative over the past five years and obtained UAV RGB and Sentinel-2 imagery. Secondly, 23 spectral features were extracted from Sentinel-2 images. Based on the UAV images, the parcel of each farmer was positioned and three topographic features of slope, aspect, and elevation were extracted. Subsequently, these 26 features were screened using the random forest algorithm and Pearson correlation analysis. Thirdly, we applied six different regression algorithms to establish fresh tea yield models for each season and evaluated their estimation accuracy. The results showed that random forest regression models were the optimal choice for estimating spring and summer yields, with the spring model achieving an R² value of 0.45, an RMSE of 40.38 kg/acre, and an rRMSE of 40.79%. Similarly, the summer model achieved an R² value of 0.5, an RMSE of 78.46 kg/acre, and an rRMSE of 39.81%. For autumn and annual yield estimation, voting regression models demonstrated superior performance, with the autumn model achieving an R² value of 0.42, an RMSE of 70.6 kg/acre, and an rRMSE of 39.77%, and the annual model attained an R² value of 0.47, an RMSE of 168.7 kg/acre, and an rRMSE of 34.62%. This study provides a promising new method for estimating fresh tea yield in different seasons at the field scale. Full article

Unmanned Aerial Vehicles (UAVs) integrated with lightweight visual cameras hold significant promise in renewable energy asset inspection and monitoring. This study presents an AI-assisted soiling detection methodology for inspecting solar photovoltaic (PV) panels, using UAV-captured RGB images. The proposed scheme introduces an autonomous end-to-end soiling detection model for common types of soiling in solar panel installations, including bird droppings and dust. Detecting soiling, particularly bird droppings, is critical due to their pronounced negative impact on power generation, primarily through hotspot formation and their resistance to natural cleaning processes such as rain. A dataset containing aerial RGB images of PV panels with dust and bird droppings is collected as a prerequisite. This study addresses the unique challenges posed by the small size and indistinct features of bird droppings in aerial imagery in contrast to relatively large-sized dust regions. To overcome these challenges, we developed a custom model, named SDS-YOLO (Soiling Detection System YOLO), which features a Convolutional Block Attention Module (CBAM) and two dedicated detection heads optimized for dust and bird droppings. The SDS-YOLO model significantly improves detection accuracy for bird droppings while maintaining robust performance for the dust class, compared with YOLOv5, YOLOv8, and YOLOv11. With the integration of CBAM, we achieved a substantial 40.2% increase in mean Average Precision (mAP50) and a 26.6% improvement in F1 score for bird droppings. Dust detection metrics also benefited from this attention-based refinement. These results underscore the CBAM’s role in improving feature extraction and reducing false positives, particularly for challenging soiling types. Additionally, the SDS-YOLO parameter count is reduced by 24%, thus enhancing its suitability for edge computing applications. Full article

Abstract: Miniaturized pixel sizes in near-eye digital displays lead to pixel emission patterns with large divergence angles, necessitating efficient beam collimation solutions to improve the light coupling efficiency. Traditional beam collimation optics, such as lenses and cavities, are wavelength-sensitive and cannot simultaneously collimate red (R), green (G), and blue (B) light. In this work, we employed inverse design optimization and finite-difference time-domain (FDTD) simulation techniques to design a collimator comprised of nano-sized photonic structures. To alleviate the challenges of the spatial incoherence nature of micro-LED emission light, we developed a strategy called dual-task optimization. Specifically, the method models light collimation as a dual task of color routing. By optimizing a color router, which routes incident light within a small angular range to different locations based on its spectrum, we simultaneously obtained a beam collimator, which can restrict the output of the light emitted from the routing destination with a small divergence angle. We further evaluated the collimation performance for spatially incoherent RGB micro-LED light in an FDTD using a multiple-dipole simulation method, and the simulation results demonstrate that our designed collimator can increase the light coupling efficiency from approximately 30% to 60% within a divergence angle of ±20° for all R/G/B light under the spatially incoherent emission. Full article

The crown area is a critical metric for evaluating tree growth and supporting various ecological and forestry analyses. This study compares three approaches, i.e., unsupervised clustering, region-based, and deep learning, to estimate the crown area of Pinus eldarica Medw. using UAV-acquired RGB imagery (2 cm ground sampling distance) and high-density point clouds (1.27 points/cm³). The first approach applied unsupervised clustering techniques, such as Mean-shift and K-means, to directly estimate crown areas, bypassing tree top detection. The second employed a region-based approach, using Template Matching and Local Maxima (LM) for tree top identification, followed by Marker-Controlled Watershed (MCW) and Seeded Region Growing for crown delineation. The third approach utilized a Convolutional Neural Network (CNN) that integrated Digital Surface Model layers with the Visible Atmospheric Resistance Index for enhanced segmentation. The results were compared against field measurements and manual digitization. The findings reveal that CNN and MCW with LM were the most effective, particularly for small and large trees, though performance decreased for medium-sized crowns. CNN provided the most accurate results overall, with a relative root mean square error (RRMSE) of 8.85%, a Nash–Sutcliffe Efficiency (NSE) of 0.97, and a bias score (BS) of 1.00. The CNN crown area estimates showed strong correlations (R² = 0.83, 0.62, and 0.94 for small, medium, and large trees, respectively) with manually digitized references. This study underscores the value of advanced CNN techniques for precise crown area and shape estimation, highlighting the need for future research to refine algorithms for improved handling of crown size variability. Full article

Variations in vegetation indices derived from multispectral images and digital terrain models from satellite imagery have been successfully used for reclamation and hazard management in former mining areas. However, low spatial resolution and the lack of sufficiently detailed information on surface morphology have restricted such studies to large sites. This study investigates the application of small, unmanned aerial vehicles (UAVs) equipped with multispectral sensors for land cover classification and vegetation monitoring. The application of UAVs bridges the gap between large-scale satellite remote sensing techniques and terrestrial surveys. Photogrammetric terrain models and orthoimages (RGB and multispectral) obtained from repeated mapping flights between November 2023 and May 2024 were combined with an ALS-based reference terrain model for object-based image classification. The collected data enabled differentiation between natural forests and areas affected by former mining activities, as well as the identification of variations in vegetation density and growth rates on former mining areas. The results confirm that small UAVs provide a versatile and efficient platform for classifying and monitoring mining areas and forested landslides. Full article

Estimating aboveground biomass (AGB) is vital for sustainable forest management and helps to understand the contributions of forests to carbon storage and emission goals. In this study, the effectiveness of plot-level AGB estimation using height and crown diameter derived from UAV-LiDAR, calibration of GEDI-L4A AGB and GEDI-L2A rh98 heights, and spectral variables derived from UAV-multispectral and RGB data were assessed. These calibrated AGB and height values and UAV-derived spectral variables were used to fit AGB estimations using a random forest (RF) regression model in Fuling District, China. Using Pearson correlation analysis, we identified 10 of the most important predictor variables in the AGB prediction model, including calibrated GEDI AGB and height, Visible Atmospherically Resistant Index green (VARIg), Red Blue Ratio Index (RBRI), Difference Vegetation Index (DVI), canopy cover (CC), Atmospherically Resistant Vegetation Index (ARVI), Red-Edge Normalized Difference Vegetation Index (NDVIre), Color Index of Vegetation (CIVI), elevation, and slope. The results showed that, in general, the second model based on calibrated AGB and height, Sentinel-2 indices, slope and elevation, and spectral variables from UAV-multispectral and RGB datasets with evaluation metric (for training: R² = 0.941 Mg/ha, RMSE = 13.514 Mg/ha, MAE = 8.136 Mg/ha) performed better than the first model with AGB prediction. The result was between 23.45 Mg/ha and 301.81 Mg/ha, and the standard error was between 0.14 Mg/ha and 10.18 Mg/ha. This hybrid approach significantly improves AGB prediction accuracy and addresses uncertainties in AGB prediction modeling. The findings provide a robust framework for enhancing forest carbon stock assessment and contribute to global-scale AGB monitoring, advancing methodologies for sustainable forest management and ecological research. Full article

Currently, multi-modal 3D object detection methods have become a key area of research in the field of autonomous driving. Fusion is an essential factor affecting performance in multi-modal object detection. However, previous methods still suffer from the inability to effectively fuse features from LiDAR and RGB images, resulting in a low utilization rate of complementary information between depth and semantic texture features. At the same time, existing methods may not adequately capture the structural information in Region of Interest (RoI) features when extracting them. Structural information plays a crucial role in RoI features. It encompasses the position, size, and orientation of objects, as well as the relative positions and spatial relationships between objects. Its absence can result in false or missed detections. To solve the above problems, we propose a multi-modal sensor fusion network, Bi-Att3DDet, which mainly consists of a Self-Attentive RoI Feature Extraction module (SARoIFE) and a Feature Bidirectional Interactive Fusion module (FBIF). Specifically, SARoIFE captures the relationship between different positions in RoI features to obtain high-quality RoI features through the self-attention mechanism. SARoIFE prepares for the fusion stage. FBIF performs bidirectional interaction between LiDAR and pseudo RoI features to make full use of the complementary information. We perform comprehensive experiments on the KITTI dataset, and our method notably demonstrates a 1.55% improvement in the hard difficulty level and a 0.19% improvement in the mean Average Precision (mAP) metric on the test dataset. Full article

In this paper, an iterative learning control scheme is applied to the interconnected RGB light strip modeled with so-called spatially interconnected systems. The proposed strategy starts with formulating a set of Kirchhoff equalities, which lead to the 2D state-space model. In the next step, the system model is transformed into its 1D equivalent using the so-called lifting approach. Subsequently, the ILC design strategy is proposed. Due to the fact that it eventually takes the form of a differential linear repetitive process, the well-established stability theory along the trial is applied. This enables the application of computationally efficient stability tests, which employ linear matrix inequalities. Such a strategy ensures a numerically tractable design procedure. The application of this strategy ensures the convergence of the output signal to the desired reference. It is important to emphasize that the considered system can be affected by disturbances, and hence, it reflects practical situations that are inevitable in engineering practice. The effectiveness of the proposed approach is illustrated with a comprehensive simulation study. Full article

Navigating crowded environments poses significant challenges for mobile robots, particularly as traditional Simultaneous Localization and Mapping (SLAM)-based methods often struggle with dynamic and unpredictable settings. This paper proposes a visual target-driven navigation method using self-attention enhanced deep reinforcement learning (DRL) to overcome these limitations. The navigation policy is developed based on the Twin-Delayed Deep Deterministic Policy Gradient (TD3) algorithm, enabling efficient obstacle avoidance and target pursuit. We utilize a single RGB-D camera with a limited field of view (FOV) for target detection and surrounding sensing, where environmental features are extracted from depth data via a convolutional neural network (CNN). A self-attention network (SAN) is employed to compensate for the limited FOV, enhancing the robot’s capability of searching for the target when it is temporarily lost. Experimental results show that our method achieves a higher success rate and shorter average target-reaching time in dynamic environments, while offering hardware simplicity, cost-effectiveness, and ease of deployment in real-world applications. Full article

Mechanochromic sensors are promising for structural health monitoring as they can visually monitor the deformation caused by discoloration. Most studies have focused on the large deformation problems over 100% strain; however, it is necessary to investigate the discoloration characteristics in a small deformation range to apply it to engineering structures, such as reinforced concrete. In this study, a photonic crystal-based discoloration sensor was investigated to determine the discoloration characteristics of the red, green, and blue (RGB) as well as hue, saturation, and value (HSV) color spaces according to displacement levels. B and S showed the highest sensitivity and linear discoloration at displacements < 1 mm, whereas R and H showed significant discoloration characteristics at displacements > 1 mm. The Vision Transformer model based on RGB and HSV channels was linearly predictable up to 4 mm displacement with an accuracy of R² 0.89, but errors were found at the initial displacement within 2 mm. Full article

Traditional limb kinematic analysis relies on manual goniometer measurements. With computer vision advancements, integrating RGB cameras can minimize manual labor. Although deep learning-based cameras aim to offer the same ease as manual goniometers, previous approaches have prioritized accuracy over efficiency and cost on PC-based devices. Nevertheless, healthcare providers require a high-performance, low-cost, camera-based tool for assessing upper and lower limb range of motion (ROM). To address this, we propose a lightweight, fast, deep learning model to estimate a human pose and utilize predicted joints for limb ROM measurement. Furthermore, the proposed model is optimized for deployment on resource-constrained edge devices, balancing accuracy and the benefits of edge computing like cost-effectiveness and localized data processing. Our model uses a compact neural network architecture with 8-bit quantized parameters for enhanced memory efficiency and reduced latency. Evaluated on various upper and lower limb tasks, it runs 4.1 times faster and is 15.5 times smaller than a state-of-the-art model, achieving satisfactory ROM measurement accuracy and agreement with a goniometer. We also conduct an experiment on a Raspberry Pi, illustrating that the method can maintain accuracy while reducing equipment and energy costs. This result indicates the potential for deployment on other edge devices and provides the flexibility to adapt to various hardware environments, depending on diverse needs and resources. Full article

Riverine environmental information includes important data to collect, and the data collection still requires personnel’s field surveys. These on-site tasks still face significant limitations (i.e., hard or danger to entry). In recent years, as one of the efficient approaches for data collection, air-vehicle-based Light Detection and Ranging technologies have already been applied in global environmental research, i.e., land cover classification (LCC) or environmental monitoring. For this study, the authors specifically focused on seven types of LCC (i.e., bamboo, tree, grass, bare ground, water, road, and clutter) that can be parameterized for flood simulation. A validated airborne LiDAR bathymetry system (ALB) and a UAV-borne green LiDAR System (GLS) were applied in this study for cross-platform analysis of LCC. Furthermore, LiDAR data were visualized using high-contrast color scales to improve the accuracy of land cover classification methods through image fusion techniques. If high-resolution aerial imagery is available, then it must be downscaled to match the resolution of low-resolution point clouds. Cross-platform data interchangeability was assessed by comparing the interchangeability, which measures the absolute difference in overall accuracy (OA) or macro-F1 by comparing the cross-platform interchangeability. It is noteworthy that relying solely on aerial photographs is inadequate for achieving precise labeling, particularly under limited sunlight conditions that can lead to misclassification. In such cases, LiDAR plays a crucial role in facilitating target recognition. All the approaches (i.e., low-resolution digital imagery, LiDAR-derived imagery and image fusion) present results of over 0.65 OA and of around 0.6 macro-F1. The authors found that the vegetation (bamboo, tree, grass) and road species have comparatively better performance compared with clutter and bare ground species. Given the stated conditions, differences in the species derived from different years (ALB from year 2017 and GLS from year 2020) are the main reason. Because the identification of clutter species includes all the items except for the relative species in this research, RGB-based features of the clutter species cannot be substituted easily because of the 3-year gap compared with other species. Derived from on-site reconstruction, the bare ground species also has a further color change between ALB and GLS that leads to decreased interchangeability. In the case of individual species, without considering seasons and platforms, image fusion can classify bamboo and trees with higher F1 scores compared to low-resolution digital imagery and LiDAR-derived imagery, which has especially proved the cross-platform interchangeability in the high vegetation types. In recent years, high-resolution photography (UAV), high-precision LiDAR measurement (ALB, GLS), and satellite imagery have been used. LiDAR measurement equipment is expensive, and measurement opportunities are limited. Based on this, it would be desirable if ALB and GLS could be continuously classified by Artificial Intelligence, and in this study, the authors investigated such data interchangeability. A unique and crucial aspect of this study is exploring the interchangeability of land cover classification models across different LiDAR platforms. Full article

In the context of escalating network adversarial challenges, effectively identifying a Webshell processed using evasion techniques such as encoding, obfuscation, and nesting remains a critical challenge in the field of cybersecurity. To address the poor detection performance of the existing Webshell detection methods for evasion samples, this study proposes a multimodal feature fusion-based evasion Webshell detection method (MMF-Detect). This method extracts RGB image features and textual vector features from two modalities: the visual and semantic modalities of Webshell file content. A multimodal feature fusion classification model was designed to classify features from both modalities to achieve Webshell detection. The multimodal feature fusion classification model consists of a text classifier based on a large language model (CodeBERT), an image classifier based on a convolutional neural network (CNN), and a decision-level feature fusion mechanism. The experimental results show that the MMF-Detect method not only demonstrated excellent performance in detecting a conventional Webshell but also achieved an accuracy of 99.47% in detecting an evasive Webshell, representing a significant improvement compared to traditional models. Full article