Xiqi Wang

Digital reconstruction of ancient buildings is very challenging due to its architectural complexity and structural delicacy. Therefore, how to apply new Earth Observation (EO) technology to digitally reconstruct the complete and real model of ancient buildings without damage has become an urgent issue. This article proposes a multi-scan point cloud hierarchical registration method for 3D reconstruction of the ancient Chinese Great Wild Goose Pagoda based on Terrestrial Laser Scanner (TLS) data. First, a multi-scan point cloud hierarchical registration method is elaborately designed for 3D reconstruction. Then, the outline feature of the ancient Chinese Great Wild Goose Pagoda is automatically extracted by point cloud slicing. Finally, fine modeling is carried out according to the outline feature extracted from the reconstructed point cloud, and after that block combination and texture mapping are conducted for 3D digitalization and documentation. The experiments indicate that the Root Mean Square Error (RMSE) of multi-scan hierarchical registration of the translation and rotation are less than 3 cm and 0.862 ^∘ , respectively, and the complete and real model of the Great Wild Goose Pagoda is obtained by fine modeling and texture mapping, which provides an important reference for 3D digital protection of ancient buildings.

Loop closure detection is a crucial technology within the field of simultaneous localization and mapping (SLAM), aiming to mitigate cumulative errors and maintain a globally consistent map. Traditional fixed-LiDAR scans have demonstrated reliable loop closure detection (LCD) performance when employing global descriptors. However, for rotary-LiDAR scans, these methods fall short in place recognition due to significant view-angle changes. To address this challenge, we propose a novel LCD algorithm tailored for rotary-LiDAR scans, named RLS-LCD. Our RLS-LCD approach utilizes submap-based recognition and a novel lightweight global descriptor to mitigate the impact of large view-angle changes inherent to rotary-LiDAR scans. Additionally, we adopt a coarse-to-fine recognition strategy to enhance recall and precision of place recognition in similar structure scenes. The experimental findings unequivocally demonstrate that the RLS-LCD algorithm not only outperforms the contemporary LCD methodologies reliant on global descriptors but also exhibits the capability to execute in real-time on computing platforms with limited computational resources.

This article proposes a two-stage simultaneous localization and mapping (SLAM) method based on using the red green blue-depth (RGB-D) camera in dynamic environments, which can not only improve tracking robustness and trajectory accuracy but also reconstruct a clean and dense static background model in dynamic environments. In the first stage, to accurately exclude the interference of features in the dynamic region from the tracking, the dynamic object mask is extracted by Mask-RCNN and optimized by using the connected component analysis method and a reference frame-based method. Then, the feature points, lines, and planes in the nondynamic object area are used to construct an optimization model to improve the tracking accuracy and robustness. After the tracking is completed, the mask is further optimized by the multiview projection method. In the second stage, to accurately obtain the pending area, which contains the dynamic object area and the newly added area in each frame, a method is proposed, which is based on a ray-casting algorithm and fully uses the result of the first stage. To extract the static region from the pending region, this paper designs divisible and indivisible regions process methods and the bounding box tracking method. Then, the extracted static regions are merged into the map using the truncated signed distance function method. Finally, the clean static background model is obtained. Our methods have been verified on public datasets and real scenes. The results show that the presented methods achieve comparable or better trajectory accuracy and the best robustness, and can construct a clean static background model in a dynamic scene.

Accurate and efficient text detection in natural scenes is a fundamental yet challenging task in computer vision, especially when dealing with arbitrarily-oriented texts. Most contemporary text detection methods are designed to identify horizontal or approximately horizontal text, which cannot satisfy practical detection requirements for various real-world images such as image streams or videos. To address this lacuna, we propose a novel method called Rotational You Only Look Once (R-YOLO), a robust real-time convolutional neural network (CNN) model to detect arbitrarily-oriented texts in natural image scenes. First, a rotated anchor box with angle information is used as the text bounding box over various orientations. Second, features of various scales are extracted from the input image to determine the probability, confidence, and inclined bounding boxes of the text. Finally, Rotational Distance Intersection over Union Non-Maximum Suppression is used to eliminate redundancy and acquire detection results with the highest accuracy. Experiments on benchmark comparison are conducted upon four popular datasets, i.e., ICDAR2015, ICDAR2013, MSRA-TD500, and ICDAR2017-MLT. The results indicate that the proposed R-YOLO method significantly outperforms state-of-the-art methods in terms of detection efficiency while maintaining high accuracy; for example, the proposed R-YOLO method achieves an F-measure of 82.3% at 62.5 fps with 720 p resolution on the ICDAR2015 dataset.

In recent years, researchers have paid increasing attention on hyperspectral image (HSI) classification using deep learning methods. To improve the accuracy and reduce the training samples, we propose a double-branch dual-attention mechanism network (DBDA) for HSI classification in this paper. Two branches are designed in DBDA to capture plenty of spectral and spatial features contained in HSI. Furthermore, a channel attention block and a spatial attention block are applied to these two branches respectively, which enables DBDA to refine and optimize the extracted feature maps. A series of experiments on four hyperspectral datasets show that the proposed framework has superior performance to the state-of-the-art algorithm, especially when the training samples are signally lacking.