ObjectiveWith the continuous advancement of industrial intelligence, intelligent robots have been widely applied in inspection scenes of key infrastructures, such as power, transportation, and energy. By performing 3D reconstruction of multi-view inspection scenes, intelligent robots can obtain accurate environmental perception capabilities, enabling autonomous operations. However, in complex inspection scenes, the existing methods often fail to extract sufficient features from weakly textured and edge regions, resulting in a low reconstruction accuracy and compromising the overall quality of the reconstruction. To address these issues, this study proposed AFE-MVSNet, a multi-view 3D reconstruction network based on adaptive feature enhancement in inspection scenes, which aims to improve the reconstruction performance in complex areas, particularly those with weak textures and edges.MethodsAFE-MVSNet consists of two main components: an adaptive feature enhancement network (AFENet) and a multi-scale depth estimation process. First, AFENet, built on a feature pyramid network, extracts multi-scale features and incorporates attention mechanisms at each upsampling stage to enhance fine details, such as edges and textures. To further improve feature representation in weakly textured and edge regions, AFENet incorporates an adaptive perception module for features (APMF) based on deformable convolutional network that dynamically adjusts kernel sampling positions and weights to enlarge the receptive field. Second, the multi-scale depth estimation network adopts a cascaded structure that refines the depth maps from coarse to fine through cost-volume construction and depth prediction at each stage. The final depth maps are fused with the corresponding color images to generate 3D colored point clouds. To enhance training, a focal loss function is used to emphasize challenging regions, thereby improving the ability of the network to learn from hard-to-extract features and the reconstruction performance.Results and DiscussionsTo validate the effectiveness of AFE-MVSNet in inspection scenes, a inspection scene dataset was constructed and a transfer learning strategy was adopted. The model was pre-trained on the DTU public dataset and then fine-tuned on the inspection scene dataset to enhance its reconstruction performance in real-world environments. AFE-MVSNet was compared with several mainstream methods on both the DTU and inspection scene datasets. The experimental results demonstrate that AFE-MVSNet significantly outperforms the existing methods in weakly textured and edge regions. It achieves an overall reconstruction error (OA) of 0.309 mm on the DTU dataset and an end-point error (EPE) of 1.006 m on the inspection scene dataset, surpassing the baseline network and its performance before fine-tuning. In addition, ablation experiments were conducted on the APMF module, attention mechanism, and focal loss function to verify the effectiveness of each component.ConclusionsTo address the poor reconstruction performance in weakly textured and edge regions of inspection scenes, this study proposed AFE-MVSNet, a multi-view 3D reconstruction network based on adaptive feature enhancement. The network aims to improve the reconstruction quality in complex areas. The main technical contributions of this study are: 1) To enhance feature representation in weakly textured and edge regions, the adaptive feature enhancement network (AFENet) was proposed, which integrates attention mechanisms to improve the extraction and representation of texture features in inspection scenes. 2) To strengthen the perceptual ability in complex areas, an adaptive perception module for features (APMF) was designed based on deformable convolutional network. This module adaptively adjusts the sampling positions and weights of the convolution kernels to enlarge the receptive field. 3) To improve the learning ability of the network, a focal loss function was introduced to enhance its ability to learn from hard-to-extract feature regions, thereby improving the reconstruction performance. 4) To improve the reconstruction ability, a inspection scene dataset was constructed. The network was first pre-trained on the DTU public dataset and then fine-tuned on the inspection scene dataset to enhance its ability to learn features. The experimental results validate the effectiveness of AFE-MVSNet in inspection scenes. The reconstructed 3D point cloud models exhibit well-preserved, weakly textured areas and clearly defined object edges. The proposed network provides a theoretical foundation for intelligent robotic applications in inspection tasks, and has significant potential for real-world engineering applications.
ObjectiveCoherent beam combining (CBC) is one of the most promising techniques for breaking through the output power limit of single channel lasers and has been one of the research hotspots in high-energy laser technology. Traditional CBC includes two schemes, tiled aperture and filled aperture schemes. Tiled aperture CBC has the ability to steer the beam by phase control, but its theoretical efficiency is relatively low in the presence of diffraction sidelobes, while filled aperture CBC can achieve high efficiency but is difficult to achieve beam self-steering. CBC based on multi-plane light conversion (MPLC) is an emerging technology that eliminates the diffraction sidelobes of tiled aperture scheme, and the pointing angle of the combined beam can be adjusted by controlling the piston phases of the sub-beams, thereby achieving efficient beam self-steering without additional optical systems such as deformable mirrors. CBC by MPLC can overcome the limitations of low energy ratio of the main lobe in tiled aperture CBC and the inability of filled aperture CBC for self-steering, and is expected to propel the development and realization of larger-scale and higher-power coherent laser sources. However, there is limited theoretical research, and no numerical analysis of the beam self-steering in filled aperture CBC systems has been reported. In this study, simulations were conducted to verify the CBC of multiple laser channels and high-dimensional mode multiplexing. Moreover, theoretical equations for beam self-steering were derived, and the steering performance under different mode bases was studied.MethodsThe principle and method of CBC by multi-plane light conversion were introduced, and the theoretical simulation model was constructed. In the implementation of CBC by MPLC, the design of phase masks at each plane is critical. The input modes of MPLC are the laser array with various phases, and the output modes are supposed to be fundamental Gaussian mode or designed to be high-order modes according to the applications. Based on the wavefront matching method, the input modes are forward propagated while the desired output modes are backward propagated, and the phase mask can be calculated and updated in an iteration process. Numerical simulations were conducted for three different application scenarios: single-mode CBC, high-dimensional mode multiplexing, and beam self-steering. Meanwhile, analytical equations of tilted Gaussian beam decomposed by orthogonal set of Laguerre?Gaussian modes were theoretically derived, demonstrating the feasibility of achieving beam self-steering in mode multiplexing MPLC by phase-only control.Results and DiscussionsSimulation results for 16-beam single-mode CBC showed that the system combining efficiency was close to 100%, and the beam quality factor was 1.03 (Fig. 3), thus achieving high-quality filled-aperture CBC. In addition, high-dimensional mode multiplexing can be achieved by phase encoding the input beam array, where different phase distributions are mapped to different high-order modes in the process of MPLC design. A simulation of 16 beams combining into 16 Hermite?Gaussian modes was carried out, with an average coupling efficiency of up to 97.4% for each mode (Fig. 5). Finally, the application of beam self-steering was analyzed, and the steering performances of three systems were compared: traditional filled-aperture CBC, single-mode CBC based on MPLC, and mode multiplexing CBC based on MPLC. As shown in Fig. 7, traditional filled-aperture CBC could not achieve beam-steering via phase-only control, single-mode CBC based on MPLC had limited beam-steering capability, while mode multiplexing CBC based on MPLC had good beam-steering capability.ConclusionsIn this work, various CBC scenarios, including single-mode combination, high-dimensional mode multiplexing, and beam self-steering, were simulated and analyzed based on the theoretical model of MPLC coherent combination. The simulation results demonstrate the simple design and excellent performance of the MPLC coherent combining system, which integrates the advantages of traditional schemes to generate high-efficiency, high-quality filled-aperture beams. Moreover, the MPLC-based CBC system can synthesize high-order modes and achieve self-steering of the combined beams through phase-only control, thus expanding the application range of filled-aperture coherent laser systems. Self-steering performance depends on high-dimensional mode multiplexing MPLC, so both the quantity and quality of multiplexed modes should be enhanced by optimizing mode mapping relationships and optical parameters. At the same time, it is necessary to improve control algorithms and accuracy to promote the realization of wide-angle beam steering.
ObjectiveShip-information monitoring is vital to marine ecological protection, fishery resource management, and sea-area safety maintenance. The current mainstream ship-monitoring methods include automatic identification systems, optical cameras, infrared thermal imaging, radar monitoring, and remote-sensing technologies. However, these technical methods present the disadvantages of active monitoring, being affected by light, limited monitoring range, susceptibility to electromagnetic-wave interference, and high cost. By contrast, distributed acoustic sensing (DAS) offers unique advantages. It can use the existing submarine communication optical cables, offers passive monitoring, resists electromagnetic interference, supports large-scale networking and long-distance monitoring, and costs lower than remote-sensing technology. Diane et al. used a submarine communication cable to obtain the direction and speed of a ship. Liu et al. used a sensitized optical cable suspended in water to obtain ship voiceprint information and proposed an array-orientation method based on adaptive phase-difference correction. These studies achieved significant advancements in ship-speed estimation, trajectory tracking, and voiceprint recognition. However, DAS has not been used to analyze ship hydrodynamic pressure field (SHPF) in ship monitoring, and the inherent parameters of ships (such as length) have not been obtained. This study proposes DAS combined with a submarine photoelectric composite cable to synchronously monitor the hydrodynamic pressure field and acoustic field of an overtopped ship. Additionally, ship information is derived by combining the signal characteristics of SHPF, acoustic-field information, a signal-enhancement algorithm, and Doppler-frequency shifts.MethodsIn this study, a phase-sensitive optical time-domain reflectometer (Φ-OTDR) combined with a submarine photoelectric composite cable was used to synchronously monitor the SHPF and acoustic field of an overtopped ship to perceive ship information. First, based on potential flow theory and the spatial response characteristics of an optical fiber, a response model of the optical fiber to SHPF was established, and the ship-overtopping time and duration were obtained based on the characteristics of the hydrodynamic pressure field. Subsequently, a signal-enhancement algorithm was used to improve the signal-to-noise ratio of the ship’s acoustic signal, and high-definition spectrum features were successfully extracted. Next, a Doppler-frequency-shift distribution model was established along the axial direction of the optical cable. Based on the least-squares fitting method, the ship speed was inverted using the experimental spectrum information, and the ship length was estimated by combining the ship-overtopping duration. This method successfully combines an existing submarine photoelectric composite cable and a Φ-OTDR to quantitatively acquire ship speed and length, thus providing a new approach for ensuring marine information security.Results and DiscussionsThe proposed method realizes SHPF monitoring, acquires the high-definition spectrum characteristics of ship spectra, and quantitatively acquires ship speed and length. The SHPF curves obtained from theoretical simulation and experiment are consistent (Fig. 9). The center of the hydrodynamic pressure curve of the ship exhibits a symmetrical distribution with a prominent negative pressure peak. Based on the characteristics of the SHPF, the ship-overtopping time and duration can be estimated. Using the obtained high-definition ship-spectrum feature map (Fig. 8), a signal with a center frequency of 21.29 Hz was selected for further processing. The ship speed was obtained via inversion using the least-squares method. The absolute error is only 0.09 m/s, and the relative error is 2.22%. The absolute error of the ship length estimated using the ship-overtopping duration is 4 m, and the relative error is 3.88%.ConclusionsIn this study, DAS and existing submarine photoelectric composite cables were used to simultaneously monitor the hydrodynamic-pressure-field and acoustic-field signals of ships. This approach is an improvement over the previous method, where DAS detects only a single physical field (acoustic field) signal of ships. Thus, it allows more ship parameters to be obtained from SHPF and acoustic-field signals. A response model of an optical fiber to an SHPF was established, and the accuracy of the theory was verified experimentally. The ship-overtopping duration was estimated using the zero-crossing negative pressure peak of the hydrodynamic pressure field. A model depicting the Doppler-frequency shift along the optical cable was established, and high-definition ship spectrum characteristics were obtained using a signal-enhancement algorithm. The ship speed was inverted using the least-squares method, and the ship length was obtained by combining the ship-overtopping duration. In this study, DAS was successfully used to synchronously detect SHPF and acoustic-field signals as well as to quantitatively acquire ship speed and length, thus expanding the information-perception ability of DAS in ship monitoring.
ObjectiveTerrestrial laser scanning (TLS) generates high-resolution three-dimensional point cloud data, yielding precise tree structures and facilitating a deeper understanding of dynamic changes within forest ecosystems. Point cloud registration is a crucial foundation for forestry applications. This study proposes a marker-free registration method for forest point clouds based on tree branching structures, which reduces the time and labor required in traditional manual registration.MethodsThis study proposes an innovative automatic registration method based on tree branching structures. First, during the preprocessing stage, ground points were filtered out, and data volume was reduced through random sampling. A graph-based connectivity method was used to separate woody points from leaf points. Subsequently, a digital terrain model (DTM) was constructed for the scene, and the bottom points were selected for density-based clustering. Two collaborative detection methods were then employed to identify tree trunks, which were used as seed points for single-tree segmentation based on comparing shortest paths. For each individual tree, key points were detected through hierarchical clustering, and key point sets were constructed. Coarse registration was performed using the four-point congruent sets (4PCS) algorithm, followed by fine registration using the point-to-plane iterative closest point (ICP) algorithm to improve registration accuracy.Results and DiscussionsTo validate the proposed method, experiments are conducted using forest point cloud data from Yeyahu National Wetland Park, Haidian Park, and the Tongji Tree dataset, which vary in terms of tree numbers and scene sizes. The Yeyahu and Haidian datasets form the experimental group, while the Tongji Tree dataset is used as an open-source reference. The experimental results show that the proposed method achieves a root mean square error (RMSE) of approximately 2.4 cm and a mean absolute error (MAE) of approximately 2.1 cm (Table 2). The registration results of this method are compared with those of the Super4PCS method, and registration overlap is used as a supplementary metric to evaluate registration performance (Table 3). Compared with other key point detection methods such as SIFT, Harris 3D, and ISS (Fig. 8), the proposed method demonstrates superior performance. It achieves an average overlap of 0.980 across five plots (see Table 4), with an average registration time of 9.74 s (Table 4). The proposed key point extraction method significantly outperforms other methods in terms of registration accuracy and efficiency.ConclusionsThis study proposes an automatic marker-free registration method for terrestrial LiDAR forest point clouds that leverage tree branching structures. The approach involves hierarchical clustering of branches to construct key points sets, followed by coarse registration using the 4PCS algorithm and fine registration of the forest point cloud scene using the point-to-plane ICP method. Experimental results indicate that using tree branch structures as key features enables high accuracy and efficiency in marker-free registration. The presence of sufficient branching features in overlapping areas is critical to ensure registration accuracy and effectiveness.
ObjectiveTraditional simultaneous localization and mapping (SLAM) back-end optimization methods typically rely on public landmarks and closed-loop detection to enhance global consistency. However, in scenarios with limited feature information or closed-loop failures, the effectiveness of optimization is often constrained. To address this issue, this study proposes a SLAM back-end optimization method based on 2D vector map constraints. The 2D vector map provides globally consistent feature point information, which can be introduced as a priori constraints in the back-end optimization process. This approach effectively compensates for the limitations of traditional methods in feature-sparse or closed-loop failure scenarios. By associating geometric features such as building contour corner points in the vector map with LiDAR sensing data, the method provides stronger global constraints for the optimization process. This enables more accurate position estimation and adjustments to the mapping results, thereby improving both the accuracy of localization and the consistency of the constructed map.MethodsBased on a graph-based optimization algorithm for laser SLAM, this study proposes an improved graph optimization algorithm that incorporates vector maps as constraints. The standard graph-based SLAM optimization algorithm was extended by incorporating vector-building contours as nodes in the factor graph and associating them with radar perception map nodes. First, a hierarchical factor graph model was constructed to divide the optimization problem into two layers: local and global. In the local optimization layer, the factor graph consisted of LiDAR odometry, IMU, vector map a priori, and frame-to-submap alignment factors. The vector map a priori factor provided additional prior information for local optimization, enhancing the local accuracy of position estimation by extracting the building contour corner points and correlating them with the corresponding LiDAR data. In the global optimization layer, the factor map further incorporated a loopback factor and a submap-to-submap alignment factor. The loopback factor was employed to correct cumulative errors when a closed loop was detected. The submap-to-submap alignment factor optimized the relative positional relationships between submaps by aligning the point-cloud data of different submaps to construct a globally consistent point-cloud map. During the optimization process, the Levenberg?Marquardt (L?M) algorithm was employed to iteratively optimize the factor map. The maximum likelihood estimation of the reference state of the point cloud map and submaps was solved for each frame by adjusting the states of the position nodes through the minimization of the error function. Through this hierarchical optimization strategy, the consistency of the global map was enhanced while maintaining the accuracy of local position estimation, resulting in a highly accurate and robust back-end optimization.Results and DiscussionsExperiments were conducted to verify the robustness, stability, and accuracy of the proposed algorithm under various conditions by designing scenarios with fewer constraints (single building) and more constraints (multiple buildings) within a campus environment. Each scenario was further categorized into loopback and non-loopback conditions. The position errors of the proposed algorithm were compared with those of Lio-Sam and Lego-Loam. The experimental results demonstrate that the accuracy of position estimation is significantly enhanced by incorporating vector map constraints. In the scenario with fewer constraints, the loopback condition exhibits a smaller position error at the starting point, which gradually accumulates as the travel distance increases. This error is corrected upon loopback. The trajectory error of the algorithm proposed in this study is slightly higher than that of Lio-Sam and Lego-Loam, indicating that the vector constraint further improves accuracy, even though the loopback mechanism is effective in reducing error (Fig. 6). The algorithm proposed in this study, under loopback conditions, reduces the RPE by 6.3% and the APE by 11.7% on average compared with Lio-Sam and reduces the RPE by 7.3% and the APE by 20.9% on average compared with Lego-Loam. In contrast, under non-loopback conditions, the proposed algorithm shows an average reduction of 8.4% in RPE and 16.3% in APE compared with Lio-Sam and an average reduction of 11% in RPE and 23.8% in APE compared with Lego-Loam (Fig. 13(a) and (b), respectively). In the scenario with more constraints, the trajectory errors of Lio-Sam and Lego-Loam under non-loopback conditions increase significantly with the travel distance, whereas the trajectories generated by the proposed algorithm are effectively corrected at corner positions owing to the constraining effect of the vector corners, thereby improving the overall accuracy (Fig. 12). The proposed algorithm, under loopback conditions, reduces the RPE by 14.1% and the APE by 13.5% on average compared with Lio-Sam, and reduces the RPE by 17.8% and the APE by 25.3% on average compared with Lego-Loam. In contrast, under non-loopback conditions, the proposed algorithm achieves an average reduction of 15.6% in RPE and 28.4% in APE compared with Lio-Sam and an average reduction of 19.9% in RPE and 36.5% in APE compared with Lego-Loam. The error reduction under multiple-building constraints is significantly greater than that in the single-building scenario with fewer constraints, indicating that the optimization effect of vector constraints becomes more pronounced with an increased number of constraints (Fig. 13(c) and (d), respectively).ConclusionsThis study proposed a SLAM back-end optimization method based on 2D vector map constraints, which significantly enhances the bit position estimation accuracy and the global consistency of the map by embedding vector corner points as strong constraints in the factor graph optimization model. Compared with traditional back-end optimization methods, the approach presented in this study demonstrates greater robustness and stability in weak-feature scenarios characterized by sparse features and closed-loop failures. Experimental results indicate that the accuracy of position estimation is significantly enhanced by incorporating vector maps as a priori constraints, and the problem of odometer drift is effectively addressed. In the loopback measurement area, the vector constraint further enhances position estimation accuracy based on the loopback mechanism. In contrast, in the non-loopback measurement area, the optimization effect of the vector constraint is more pronounced, effectively reducing error accumulation. The method presented in this study offers a new technical approach for high-precision laser map construction and provides a reliable solution for SLAM system optimization in weak-feature scenarios. Future research may further explore the application of this method to map updating, using optimized map construction results to supplement the vector map with local details, thereby enhancing the richness and practical utility of the map.