Catálogo de publicaciones - revistas

<jats:title>Abstract</jats:title><jats:p>Neural radiance fields (NeRF) offer the potential to benefit 3D reconstruction tasks, including aerial photogrammetry. However, the scalability and accuracy of the inferred geometry are not well‐documented for large‐scale aerial assets. We aim to provide a thorough assessment of NeRF in 3D reconstruction from aerial images and compare it with three traditional multi‐view stereo (MVS) pipelines. However, typical NeRF approaches are not designed for large‐format aerial images, which result in very high memory consumption (often cost‐prohibitive) and slow convergence when directly applied to aerial assets. Despite a few NeRF variants adopting a representation tiling scheme to increase scalability, the random ray‐sampling strategy during training still hinders its general applicability for aerial assets. To perform an effective evaluation, we propose a new scheme to scale NeRF. In addition to representation tiling, we introduce a location‐specific sampling technique as well as a multi‐camera tiling (MCT) strategy to reduce memory consumption during image loading for RAM, representation training for GPU memory and increase the convergence rate within tiles. The MCT method decomposes a large‐frame image into multiple tiled images with different camera models, allowing these small‐frame images to be fed into the training process as needed for specific locations without a loss of accuracy. This enables NeRF approaches to be applied to aerial datasets on affordable computing devices, such as regular workstations. The proposed adaptation can be implemented to adapt for scaling any existing NeRF methods. Therefore, in this paper, instead of comparing accuracy performance against different NeRF variants, we implement our method based on a representative approach, Mip‐NeRF, and compare it against three traditional photogrammetric MVS pipelines on a typical aerial dataset against lidar reference data to assess NeRF's performance. Both qualitative and quantitative results suggest that the proposed NeRF approach produces better completeness and object details than traditional approaches, although as of now, it still falls short in terms of accuracy. The codes and datasets are made publicly available at <jats:ext-link xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="https://github.com/GDAOSU/MCT_NERF">https://github.com/GDAOSU/MCT_NERF</jats:ext-link>.</jats:p>

<jats:title>Abstract</jats:title><jats:p>Most spectral–spatial classification methods for hyperspectral images (HSIs) can achieve satisfactory classification results. However, the common problem faced with these approaches is the need for a long training time and sufficient training samples. To address this issue, this study proposes an effective spectral–spatial HSI classification method based on superpixel merging, superpixel smoothing and broad learning system (SMS‐BLS). The newly introduced parameter‐free superpixel merging technique based on local modularity not only enhances the role of local spatial information in classification, but also maintains class boundary information as much as possible. In addition, the spectral and spatial information of HSIs is further fused during the superpixel smoothing process. As a result, with limited training samples, using merged and smoothed superpixels instead of pixels as input to the broad learning system significantly improves its classification performance. Moreover, the merged superpixels weaken the dependence of the classification results on the superpixel segmentation scale. The effectiveness of the proposed method was validated on three HSI benchmarks, namely Indian Pines, Pavia University and Salinas. Experimental and comparative results show the superiority of the method to other state‐of‐the‐art approaches in terms of overall accuracy and running time.</jats:p>

<jats:title>Abstract</jats:title><jats:p>Simultaneous localization and mapping (SLAM) is a very challenging yet fundamental problem in the field of robotics and photogrammetry, and it is also a prerequisite for intelligent perception of unmanned systems. In recent years, 3D LiDAR SLAM technology has made remarkable progress. However, to the best of our knowledge, almost all existing surveys focus on visual SLAM methods. To bridge the gap, this paper provides a comprehensive review that summarizes the scientific connotation, key difficulties, research status, and future trends of 3D LiDAR SLAM, aiming to give readers a better understanding of LiDAR SLAM technology, thereby inspiring future research. Specifically, it summarizes the contents and characteristics of the main steps of LiDAR SLAM, introduces the key difficulties it faces, and gives the relationship with existing reviews; it provides an overview of current research hotspots, including LiDAR‐only methods and multi‐sensor fusion methods, and gives milestone algorithms and open‐source tools in each category; it summarizes common datasets, evaluation metrics and representative commercial SLAM solutions, and provides the evaluation results of mainstream methods on public datasets; it looks forward to the development trend of LiDAR SLAM, and considers the preliminary ideas of multi‐modal SLAM, event SLAM, and quantum SLAM.</jats:p>

<jats:title>Abstract</jats:title><jats:p>Land cover change detection (LCCD) with remote sensing images (RSIs) is important for observing the land cover change of the Earth's surface. Considering the insufficient performance of the traditional self‐attention mechanism used in a neural network to smoothen the noise of LCCD with RSIs, in this study a novel cross‐attention neural network (CANN) was proposed for improving the performance of LCCD with RSIs. In the proposed CANN, a cross‐attention mechanism was achieved by employing another temporal image to enhance attention performance and improve detection accuracies. First, a feature difference module was embedded in the backbone of the proposed CANN to generate a change magnitude image and guide the learning progress. A self‐attention module based on the cross‐attention mechanism was then proposed and embedded in the encoder of the proposed network to make the network pay attention to the changed area. Finally, the encoded features were decoded to obtain binary change detection with the ArgMax function. Compared with five methods, the experimental results based on six pairs of real RSIs well demonstrated the feasibility and superiority of the proposed network for achieving LCCD with RSIs. For example, the improvement for overall accuracy for the six pairs of real RSIs improved by our proposed approach is about 0.72–2.56%.</jats:p>

<jats:title>Abstract</jats:title><jats:p>The precise geometric representation and ability to handle long‐tail targets have led to the increasing attention towards vision‐centric 3D occupancy prediction, which models the real world as a voxel‐wise model solely through visual inputs. Despite some notable achievements in this field, many prior or concurrent approaches simply adapt existing spatial cross‐attention (SCA) as their 2D–3D transformation module, which may lead to informative coupling or compromise the global receptive field along the height dimension. To overcome these limitations, we propose a hierarchical occupancy (HierOcc) network featuring our innovative height‐aware cross‐attention (HACA) and hierarchical self‐attention (HSA) as its core modules to achieve enhanced precision and completeness in 3D occupancy prediction. The former module enables 2D–3D transformation, while the latter promotes voxels’ intercommunication. The key insight behind both modules is our multi‐height attention mechanism which ensures each attention head corresponds explicitly to a specific height, thereby decoupling height information while maintaining global attention across the height dimension. Extensive experiments show that our method brings significant improvements compared to baseline and surpasses all concurrent methods, demonstrating its superiority.</jats:p>

<jats:title>Abstract</jats:title><jats:p>Fine‐grained indoor navigation services require obstacle‐level indoor maps to support, but since indoor environments are affected by human activities, resulting in frequent changes in indoor spatial layouts, and indoor environments are easily affected by light and occlusion, the vast majority of indoor maps are at room level, limiting indoor obstacle‐level navigation path planning. To solve this problem, this paper proposes a hierarchy relation graph (HRG) construction method based on RGB‐D. Firstly, the semantic information extraction of indoor scenes and elements is realized by the output transformed PSPNet and YOLO V8 models, and the bounding box of each element is obtained based on YOLO V8. Then an algorithm for determining the hierarchical relationship of indoor elements is proposed, which calculates the correlation between the two elements from both plane and depth dimensions and constructs a HRG of indoor elements based on directed trees. Finally, comparative experiments are designed to validate the proposed method. Experiments showed that the proposed method can construct HRGs in a variety of scenes; the hierarchy relation detection rate is 88.28%; the accuracy of hierarchy relation determination is 73.44%; and the single‐scene HRG can be generated in 3.81 s.</jats:p>

<jats:title>Abstract</jats:title><jats:p>Forest topographic survey is a problem that photogrammetry has not solved for a long time. Forest canopy height is a crucial forest biophysical parameter which is used to derive essential information about forest ecosystems. In order to construct a canopy height model in forest areas, this study extracts spectral feature factors from digital orthophoto map and geometric feature factors from digital surface model, which are generated through aerial photogrammetry and LiDAR (light detection and ranging). The maximum information coefficient, Pearson, Kendall, Spearman correlation coefficients, and a new proposed index of relative importance are employed to assess the correlation between each feature factor and forest vertical heights. Gradient boosting decision tree regression is introduced and utilised to construct a canopy height model, which enables the prediction of unknown canopy height in forest areas. Two additional machine learning techniques, namely random forest regression and support vector machine regression, are employed to construct canopy height model for comparative analysis. The data sets from two study areas have been processed for model training and prediction, yielding encouraging experimental results that demonstrate the potential of canopy height model to achieve prediction accuracies of 0.3 m in forested areas with 50% vegetation coverage and 0.8 m in areas with 99% vegetation coverage, even when only a mere 10% of the available data sets are selected as model training data. The above approaches present techniques for modelling canopy height in forested areas with varying conditions, which have been shown to be both feasible and reliable.</jats:p>