Abstract:Ship exhaust gas testing is of great significance for the implementation of environmental protection regulations and the promotion of sustainable development. At present, the shipping industry is developing in a green and carbon-free direction, and the unique advantages of clean fuels in environmental protection are gradually becoming the mainstream of ship fuel transformation, but the use of clean fuels has caused some new emission problems, and new demand for exhaust gas detection is emerging, which in turn affects the applicability of existing testing methods and equipment. By analyzing the emissions of traditional fossil fuels and five new clean fuels, this paper proposes an emission detection list and points out the components that need to be tested outside the regulations. According to the location of exhaust gas detection, the detection methods of flue gas and plume were introduced, and the shortcomings of plume detection in unconventional emissions were pointed out. Eight commonly used gas analyzers were introduced, and the current gas analyzers were analyzed and evaluated from five dimensions: volume, applicable detection methods, types of components, detection time and accuracy, and finally the development direction of exhaust gas detection methods and equipment was prospected, and it was pointed out that miniaturization, intelligence and low power consumption will be the focus of the subsequent development of testing equipment. The review and research of existing ship exhaust gas detection methods will help guide the research and development of related testing equipment.

Abstract:Surface electromyography directly reflects muscle activity, effectively capturing muscle contraction patterns and intensity, making it widely used in gesture recognition. However, it′s sparsity, non-linearity, and noise interference pose significant challenges for feature extraction. To address this, we propose the RASTNet model, using ResNet50 as the backbone and replacing the 3×3 convolution in each layer′s last block with an Atrous Spatial Pyramid Pooling module to capture multi-scale information via dilated convolutions. An STConv module, incorporating a triple attention mechanism into SCConv, is added to enhance the fusion of channel and spatial features. Experiments on the NinaPro DB1 and DB5 datasets, augmented with four methods, show that RASTNet improves accuracy by 1.83% and 1.57% on average. Compared with models like ResNeXt, Swin Transformer, and CnovNeXt under simulated noise, RASTNet outperforms in recall rate, F1 score, and other metrics. It also remains superior to the latest closed-source models without noise, demonstrating robustness and noise resistance in complex gesture recognition tasks. Additionally, RASTNet shows strong generalization across datasets, enhancing its real-world applicability and robustness.

Abstract:This paper is based on the resonance phenomenon of a square flat no leads package device when it is applied in Ku band, which leads to circuit insertion loss of more than 3dB and serious deterioration of circuit flatness. Based on this phenomenon, circuit resonance analysis is carried out and solutions are explored. In this paper, the RF characteristics and equivalent circuit of coplanar waveguide with three types of defected ground structures are theoretically analyzed. Based on the coplanar waveguide theory, the structural parameters of the third type of defected ground structure and the characteristics of the transmission line with ground cavities at both ends of the transmission line are simulated and analyzed by CST simulation software. The research shows that the RF output pin of QFN package device is easy to form a defective coplanar waveguide with gaps with the surrounding pins and the surrounding ground, which makes the circuit equivalent to an LC parallel resonant circuit. In addition, the length of cavity l, the width of cavity w, the width of gap g, the distance from the gap to the side wall of the cavity d and the simultaneous presence of cavities at both ends of the transmission line will affect the characteristics of the transmission line, making the circuit prone to resonant points in the low frequency band. In this paper, the circuit is optimized and the resonance phenomenon is effectively eliminated by removing the gap g in the equivalent defect ground coplanar waveguide structure in the circuit. Then the method of eliminating resonance phenomenon and circuit design method adopted through theoretical and simulation analysis can effectively improve the flatness in the microwave band. This method can be widely used for reference in the design of modern high-density integrated fourth generation microwave circuits.

Abstract:Aiming at the problems that the traditional osprey algorithm has low convergence efficiency and is prone to fall into local optimality when solving the path planning problem of intelligent agents, an improved osprey algorithm is proposed. This algorithm integrates the Tent chaotic mapping to enhance the diversity of the population. Secondly, a weight factor and a Gaussian mutation strategy are introduced to prevent the algorithm from falling into local optimality, effectively improving the global search ability. To verify the effectiveness of this algorithm, 10 standard test functions and 2 sets of grid environments with different complexities are selected for experiments. The results show that the improved osprey optimization algorithm has good convergence and convergence rate on the standard test functions. Moreover, compared with the traditional osprey algorithm, the average value of the path optimization length of the improved osprey algorithm decreases by 9.08% and the standard deviation decreases by 49.18% in Environment 1, and the average value of the path optimization length decreases by 6.51% and the standard deviation decreases by 39.62% in Environment 2, which reflects a better path optimization effect and stability.

Abstract:To address the issues of poor human-robot interaction compliance caused by fixed impedance parameters and insufficient dynamic disturbance compensation in traditional upper limb rehabilitation robots, this study proposes an impedance model-based interactive control strategy. First, an adaptive robust controller was designed to compensate for system uncertainties including model parameters and external disturbances. Second, an adaptive impedance parameter regulator was developed to overcome the compliance deficiency induced by fixed impedance parameters, which dynamically adjusts impedance parameters by establishing correlations between patient exertion levels, robot motion states, and impedance parameters. Trajectory tracking simulation results demonstrate that compared with conventional PD control, the proposed adaptive robust control reduces NRMSD of shoulder and elbow joint trajectories by 35.20% and 63.31%, respectively, under dynamic disturbance conditions. Active compliance simulations reveal that the system can dynamically adjust training trajectories based on patient exertion levels and track them effectively. Compared with variable impedance control using PD compensation, the proposed adaptive robust compensation-based variable impedance control achieves 70.79% and 54.53% reductions in shoulder and elbow joint NRMSD, respectively. These results indicate that the proposed control scheme not only enhances compliance but also exhibits superior robustness, effectively meeting the requirements for patient rehabilitation training.

Abstract:An improved YOLOv11 real-time object detection algorithm is proposed to address the detection difficulties and inaccurate positioning of small and occluded targets in complex dynamic scenarios of intelligent vehicles. Firstly, in response to the difficulty of small object recognition caused by the loss of pooling layer features in the backbone network, DSEAIFI was proposed on the basis of AIFI to replace the pooling layer in the backbone network. Secondly, in order to improve the neck network′s ability to utilize and fuse features, as well as enhance its ability to detect occluded targets, the MFFNeck network was proposed, which improved the model′s ability and adaptability to fuse contextual features. Finally, in order to further improve the adaptability of the network to complex dynamic environments and highlight the importance level of advanced features in the feature map, a LAAFPN network designed for the detection head was integrated into the head network. To verify the performance of the proposed algorithm, simulations and real vehicle experiments were conducted, and the simulation results showed that the improved algorithm was effective on the KITTI dataset mAP@0.5 and mAP@0.5 0.95 is 91.1% and 70.1% respectively, which is an improvement of 2.1% and 3.8% compared to the basic model. The actual vehicle experiment results show that the average detection accuracy of the proposed algorithm is 92.7%, which is 4.3% higher than the basic model and has good real-time performance.

Abstract:Aiming at the high-efficiency, high-resolution and non-contact measurement requirements for impurities, holes, cracks and imperfections in industrial production processes such as additive printing, aeroengine blades and complex components, a wideband laser ultrasonic signal modal separation method is proposed. Through the effective combination of array signal time-domain average denoising method, improved empirical mode decomposition algorithm, wavelet denoising method for wideband signal multi feature analysis, and variational mode decomposition algorithm, the denoising, feature enhancement and modal separation of laser ultrasonic signals are achieved. Traditional ultrasound B-scan, C-scan imaging, and synthetic aperture focusing imaging algorithms are used to achieve high-precision two-dimensional imaging of defects is achieved. Based on forming a three-dimensional matrix through spatiotemporal dynamic scanning, three-dimensional quantitative display of defects is achieved. Two sets of laser ultrasonic defect measurement systems with high precision scanning type for five-axis machine tools and free scanning type for robot arms are developed. High-precision ultrasonic array sensors based on optical interference and electromagnetic ultrasonic sensor are designed. Four preset defect simulation test blocks are designed, including flat bottom holes, transverse through holes, surface cracks, and internal cracks, for preliminary verification. The actual blade and turbine disk defect detection results show that the system can effectively detect defects with a size of 0.1 mm, with a defect size error of less than 10% and a position error of less than 0.3 mm. The maturity of the developed laser ultrasonic detection system has reached level 6, and it has further promotion and application value in fields such as aerospace, navigation, nuclear power, rail transportation, pressure vessels and pipelines, toxic gas containers and pipelines, etc.

Abstract:The traditional method of line selection for single-phase grounding fault in small current grounding system usually adopts the line selection model based on one-dimensional signal, which has some problems such as low line selection accuracy and weak noise resistance. Based on the above problems, this paper proposes a line selection method for single-phase grounding fault of distribution network small current grounding system based on optimized VMD and dual-channel CNN-MATT: frost and ice algorithm is used to optimize the decomposition layer number and penalty factor of VMD, and fuzzy entropy algorithm is used to select the IMFs component with minimum fuzzy entropy as the noise reduction output signal. Gram-angle field algorithm is used to transform the denoised signal into two-dimensional spatial domain image, and the fault database is constructed. The GASF and GADF images are taken as the input of dual-channel neural network, and the fault features contained in the images are extracted and learned by CNN-MATT, and the fault lines are selected. In order to verify the effectiveness of the proposed method, MATLAB/Simulink and RTLAB closed-loop simulation platform are used in this paper, and the proposed model is compared with three kinds of line selection models under the premise of adding noise. The experimental results show that the accuracy rate of the proposed algorithm is as high as 99.4%, and it can maintain an accuracy of more than 95% under different noise conditions, which is superior to the other three line selection models, and overcomes the problems of low accuracy and poor noise resistance of traditional fault line selection methods.

Abstract:Aiming at the problems of insufficient spatio-temporal feature fusion and failure to fully utilize the rich skeleton information in the existing action recognition algorithms, this paper proposes a dual-stream fusion action recognition model based on cross-modal synergetic perception. Firstly, this paper proposes a dual-stream fusion model, which obtains the global information of the two modules by fusing the RGB video stream and the skeleton stream, realizing the complementary advantages; proposes a spatio-temporal interaction and attention enhancement module, which realizes the in-depth synergistic and dynamic complementarity of spatio-temporal features and dynamically enhances the attention weight of the relevant spatio-temporal region; and finally, designs a Multimodal Feature Fusion Module.Feature Fusion Module, which will be enhanced by feature fusion through the outputs of RGB video streams and skeleton streams, and fully exploits the complementary information between RGB visual appearance and human skeleton motion through adaptive weight assignment and cross-modal interaction, so as to improve the accuracy of action recognition. The results of multiple sets of experiments show that this CC-DFARM achieves high accuracy on the NTU RGB+D and NTU RGB+D 120 datasets of action recognition, obtaining 97.2% and 92.3% accuracies, respectively, and improving the accuracy by 3.6% and 3.2% compared to the baseline method MMTM. The results show that the model can fully extract and utilize the human skeleton information, and at the same time fully integrate the spatio-temporal features to improve the accuracy of action recognition.

Abstract:A method for diagnosing weak defects in cables based on the difference enhancement of the reflection coefficient spectra is proposed because it is difficult to identify weak defects using the current cable inspection techniques due to the slight difference between their reflection coefficient spectra and intact cables. First, the internal link between the reflection coefficient spectrum and the defect degree is explained, and the mathematical model of the cable line′s reflection coefficient spectrum is constructed. Second, an analysis is conducted to compare the reflection coefficient spectrums of the complete cable and the weak defect. A weak defect location function based on differential spectrum enhancement is devised with the goal of identifying the distinguishable difference between the two, and a reflection coefficient spectrum recovery approach is suggested. Using the current detection data, this method can recreate the theoretical reflection coefficient spectrum of the cable in good condition. Defect characteristics are successfully improved, and the sensitivity and accuracy of defect identification are raised, by processing the difference between the restored intact reflection coefficient spectrum and the reflection coefficient spectrum that contains defects. Ultimately, simulation and field tests were used to confirm the method′s efficacy. The findings demonstrated that the technique could precisely detect weak cable flaws and increase diagnostic accuracy, with a localization error of less than 4%.

Abstract:In order to solve the problem of low accuracy of deep learning network when the training sample size is insufficient, a specific radiation source recognition method based on time-frequency data enhancement and contrast learning is proposed. Firstly, I/Q information is extracted from the radiation source signal, and multi-modal information is constructed through continuous wavelet transform and time-frequency data of Welch power spectrum enhancement. In this way, the small sample data is expanded and sent into contrast learning networks for feature extraction. In addition, a weighting and loss function based on cross entropy and supervised contrast loss is designed. The features of specific radiation source signals are fully extracted to ensure that the two feature vector distributions have the consistency of cosine loss, and the optimal model is saved after training. Finally, part of the training set data is used to fine-tune the model. The proposed approach was evaluated on ADS-B dataset and WiFi dataset, compared with baseline models, and compared with 28 data enhancement combinations. Experimental results show that the method proposed in this paper achieves better results than the existing methods, and the data enhancement combination method proposed in this paper has the best effect. Specifically, when the ratio of the number of labeled training samples to the number of all training samples is 5%, the recognition accuracy on the ADS-B dataset is 87.30%, which is 6% higher than that of the baseline model. The recognition accuracy on WiFi data set is 94.07%, which is 55.39% higher than the baseline model.

Abstract:To improve the accuracy of online fault prediction for traction transformers, an online fault diagnosis method based on random forest feature optimization and improved honey badger optimization algorithm is proposed. Firstly, the SMOTE algorithm is used for data balancing processing, and the uncoded ratio method is adopted to expand the fault diagnosis; secondly, the feature vector set is ranked by importance using RF, and then input into the Extreme Learning Machine, Support Vector Machine, and Long Short Term Memory Neural Network to obtain the optimal combination of the base model and the number of features; then, the honey badger optimization algorithm was improved by combining Tent chaotic mapping strategy, improved control factor, and pinhole imaging strategy, and compared with other optimization algorithms to demonstrate its effectiveness in optimization ability, stability, and convergence speed; finally, by combining the improved honey badger optimization algorithm with the optimal base model and number of features, the problem of hyperparameter setting in the base model was effectively solved. The experimental results show that the fault diagnosis accuracy of the proposed online fault prediction model for traction transformers is 96.05%, which is 2.44% higher than that of HBA-LSTM, verifying the effectiveness of the proposed method.

Abstract:Aiming at the problems of complicated control structure and slow dynamic response speed of three-phase PWM rectifier direct power control under the double-loop control structure, and considering the applicability of control algorithms applied in microprocessors, a single-loop discrete sliding-mode direct power control method based on the discrete nonlinear disturbance observer for three-phase PWM rectifier is proposed. Firstly, the rectifier DC-side power model and the grid-side active power model are considered as a whole to obtain the DC-side-grid-side active power model. Second, a discrete nonlinear disturbance observer is designed and a novel sliding mode surface containing power perturbation estimation is constructed to improve the power perturbation resistance of single-loop discrete sliding mode control and reduce its power tracking error. Finally, the dual-loop PI control, the traditional discrete sliding mode control and the control proposed in this paper are compared on a processor-in-the-loop testbed. Compared with the dual-loop PI control, the regulation time and voltage dips of the proposed control are reduced by 90%, 35.7% and 91%, 41% for resistive load surge and constant power load surge, respectively. During steady state operation, the total harmonic distortion rate of the current of the proposed control is reduced by 36% and 15% compared with that of the traditional discrete sliding mode control and dual-loop PI control, respectively. The test results verify the effectiveness and superiority of the proposed control method.

Abstract:To address the challenges of low recognition accuracy for small target defects and insufficient multi-scale feature capture capability in traditional vision-based methods for solar cell inspection, this study proposes an improved YOLOv8 algorithm based on cross-scale feature enhancement and dynamic parameter optimization. First, a multi-branch residual structure is designed as the core, integrating re-parameterization techniques and adjustable dilated convolution to construct the dilation-enhanced reparametrized residual block, which enhances contextual awareness of target defects through cross-layer feature interaction, thereby improving detection accuracy. Second, deformable convolutional networks version 2 are embedded into the C2f module, combined with an auxiliary detection module to build a dynamic feature adaptation network, improving geometric feature extraction for micro-defects. Finally, a wise intersection over union loss function with a dynamic focusing mechanism is introduced to optimize bounding box matching and enhance regression precision. Experimental results demonstrate that the improved model achieves a mean average precision of 91.8% with only 3.03 M parameters, outperforming the baseline model by 4% while maintaining lightweight architecture and improved detection performance.

Abstract:To address the issues of missed detections and low accuracy caused by high occlusion and large scale variations in dense pedestrian detection, this paper proposes an efficient improved RT-DETR algorithm, RSH-RTDETR, for complex scene dense pedestrian detection. Firstly, the Regocn module is proposed to improve the backbone, using limited ordinary convolutions for feature extraction, followed by a linear transformation operation. Meanwhile, RepConv is used on the gradient flow branch to compensate for the performance loss caused by discarding residual blocks and enhance the feature extraction and gradient flow capabilities, achieving better detection of targets of different scales while reducing the computational load and parameter count. Secondly, a 160×160 S2 detection layer is introduced in the neck to enhance the detection ability of small-scale pedestrian targets during the feature fusion stage. Finally, the Haar wavelet downsampling module (HWD) is adopted to expand the receptive field, reduce model complexity, and improve the detection accuracy of occluded pedestrian targets. Ablation and comparison experiments were conducted on the CrowdHuman dataset, achieving an mAP50 of 86.6% and an mAP50.95 of 57.8%. Compared with the original algorithm, mAP50 was improved by 1.2% and mAP50.95 by 1.9%, with a 40% reduction in parameters. It also outperformed the RT-DETR algorithm on the Wider person dataset. Experimental results show that RSH-RTDETR improves the accuracy of dense pedestrian detection while reducing the parameter count compared to the RTDETR-R18 model, and outperforms other algorithms. The improved algorithm in this paper achieves lightweight while maintaining high accuracy, demonstrating excellent performance in dense pedestrian detection tasks in complex scenes.

Abstract:Aiming at the problem of large trajectory error and map drift of most slam algorithms in outdoor long-distance environment, a SLAM algorithm based on IEKF tightly coupled lidar and IMU is proposed, and a globally consistent laser 3D point cloud map is constructed. Firstly, the IMU state model is constructed and the state is estimated by forward propagation, and the back propagation is used to compensate the motion of the point cloud, and then the IMU data and radar data are fused by iterative extended Kalman filter to obtain the front-end laser odometer; the loop back detection module is introduced to construct the triangle descriptor in the point cloud and match the edges of the triangle descriptor to achieve closedloop detection; finally, in the back-end optimization part, GTSAM is used to build a factor map, which integrates IMU pre integration factor, odometer factor and loop detection factor to eliminate cumulative errors, improve positioning accuracy and reduce map drift. Experiments show that compared with FAST-LIO2 algorithm, the APE RMSE of the proposed algorithm in KITTI data set and self collected data set is reduced by 50.06% and 33.65%, respectively, and the drift on the z axis is reduced, which can build a closed dense point cloud map.

Abstract:Semantic segmentation is a key technology in autonomous driving. Outdoor scene image semantic segmentation faces challenges like environmental complexity and sample imbalance, leading to suboptimal performance. To address these issues, this paper proposes a semantic segmentation network for outdoor scenes based on feature branch enhancement, FBE-Net. FBE-Net adopts an encoder-decoder architecture and designs a feature enhancement branch. It utilizes multi-scale dilated attention to capture key features and enhance overall accuracy, and employs a memory module to address sample imbalance. Simultaneously considering lightweight design. We collected campus scene data using an HD camera, annotated it with semantic labels, and created a campus scene semantic segmentation dataset. Experiments were conducted on the Cityscapes dataset and the self-built dataset. The experimental results showed that FBE-Net achieved a mIoU of 79.64% on the Cityscapes dataset and 78.01% on the self-made dataset, outperforming mainstream semantic segmentation methods.

Abstract:To improve high-performance multi-scale object detection, particularly the accuracy of small object detection, and reduce the probability of traffic accidents, this study proposes an enhanced YOLO11 model with a multi-scale context-enhanced attention mechanism for vehicle detection. Firstly, the RPCSPELAN5 structure is designed and introduced in the backbone network to replace the C3k2 module, enhancing feature extraction capability and information aggregation. Secondly, a DSM module is created and added to the neck network, which incorporates a dynamic upsampling mechanism and a simple, parameter-free attention mechanism to improve feature fusion for small objects. Finally, the neck network is further improved by adopting a Haar wavelet-based downsampling module, which enhances semantic segmentation performance and contextual continuity. Experiments on the VOC2012 and COCO datasets demonstrate significant improvements across multiple evaluation metrics. On the VOC2012 dataset, the improvements in P, R, mAP50, and mAP50.95 were 0.2%, 5.3%, 3.4% and 4.2%, respectively. On the COCO dataset, the improvements were 7.7%, 6.0%, 8.7% and 6.5%, respectively. The proposed algorithm exhibits superior performance in multi-scale object detection, particularly in small object detection accuracy, effectively enhancing vehicle detection precision and contributing to the reduction of traffic accidents.

Abstract:In response to the demand for improving the accuracy of existing backscatter imaging algorithms and the limitations of noise resistance, this paper proposes an inverse scattering imaging method based on the integration of Subspace Optimization Method with Variance Reduction and the M2Net. Within the optimized SOM framework, the Stochastic Variance Reduced Gradient method is introduced, employing a two-layer loop structure that randomly samples a small subset of data in each iteration to update the model, thereby reducing variance through correction terms and improving computational efficiency. Building upon this, a U-shaped nested model called M2Net, incorporating M-shaped residual blocks with multi-scale layers, is constructed. The initial reconstruction results are used as input data for the deep network training of M2Net, enabling further high-precision reconstruction of the scatterer structure. Compared with traditional methods, this method improves structural similarity by 10%~30% and reduces root mean square error by 5%~15%, indicating excellent noise resistance performance and the ability to achieve high-precision image reconstruction.

Abstract:A lightweight detection model YOLO-DAS was proposed to solve the problems of small target size, complex background interference and low efficiency of multi-scale feature fusion in UAV aerial images. A dynamic multi-scale sensing convolution module DMSConv is constructed to enhance the feature capture capability. The context-aware feature recombination upsampling ADEPT was designed to optimize the feature map reconstruction process to improve the integration accuracy of context information. The neck network is reconstructed using the bidirectional global-local spatial attention SCOPE, and the single path fusion limitation is broken through the bidirectional feature interaction. A shallow small target detection layer is added to strengthen the localization information extraction of low-level features. Based on the VisDrone2019 dataset, the model achieved 39.8% and 23.7% in mAP0.5 and MAP0.5:0.95 indexes, respectively, which increased by 8.4% and 5.1% compared with the benchmark YOLOv8n. The accuracy and recall rate increased by 8.1% and 7% simultaneously, and the number of parameters decreased by 0.49 M. It provides an effective solution for small and medium-sized target detection in UAV aerial images.

Abstract:With the wide application of deep learning technology in rail inspection, visual inspection methods in the field of rail fasteners have been increasingly studied. Aiming at the efficiency bottleneck of constructing defective samples in the current rail fastener data set, and the relative lack of means to detect loose parts based on image data, this paper proposes a rail fastener detection method based on data enhancement and YOLO model. In this study, the line array camera mounted on the inspection vehicle collects images to obtain raw data and texture information, uses the a priori information of the image to control the point cloud data to efficiently generate mask images and label files containing contour information, and realizes the migration and fusion of texture information based on the style migration model. Aiming at the demand of synchronization based on image data to realize the detection of missing and other states and loose states, the attention mechanism and adaptive splicing layer are introduced, and the multi-task detection model is constructed to realize the rapid identification of fastener states and the accurate segmentation of the bolt region, and the average accuracy of target detection reaches 92.14%, and the pixel accuracy of semantic segmentation reaches 89.6%. The method in this paper effectively improves the efficiency of data enhancement and reduces the probability of leakage detection for bolt states in the field of 2D images.

Abstract:The surface quality of steel plate products significantly impacts their performance and market competitiveness. To address the challenges of insufficient detection accuracy, frequent false positives, and severe missed detections in steel plate surface defect detection, this paper proposes an improved model, SGF-YOLOv8n, based on YOLOv8n. First, a Slim-neck structure is introduced to effectively reduce the model′s parameter count and computational complexity, thereby enhancing computational efficiency. Second, the GAM attention mechanism is integrated to strengthen the model′s perception of global features, improving its ability to detect subtle defects. Finally, the Focaler-IoU loss function is employed to further optimize the model′s localization accuracy when dealing with blurry boundaries and small defect areas. Additionally, to address the issue of limited dataset samples, data augmentation techniques were applied to expand the NEU-DET dataset, followed by extensive experiments. The experimental results demonstrate that SGF-YOLOv8n achieves an mAP50 of 81.6% on the NEU-DET dataset, representing a 3.8% improvement over the baseline model. Furthermore, in generalization experiments on the GC10-DET dataset, SGF-YOLOv8n achieved an mAP50 of 70.4%, a 6.7% increase compared to the baseline. These results indicate that the proposed algorithm exhibits robust performance and high effectiveness.

Abstract:This paper proposes a detection technology of Dual-Plane Linear Array Electromagnetic Tomography, aimed at enhancing the traffic management and congestion control capabilities of traffic authorities, especially for traffic accident detection on mountain highways. Considering the characteristic of the road sides being extended, the paper designs a sensor structure composed of two parallel planar coil arrays, each linearly arranged with five coils. A three-dimensional finite element method is employed to simulate the feasibility of Dual-Plane Linear Array Electromagnetic Tomography for detecting traffic accidents and to analyze the distribution characteristics of the electromagnetic field in the object field as well as the characteristics of the sensitivity maps. The paper studies the effect in reconstructing vehicle distribution using LBP algorithm, Tikhonov Regularization algorithm, and Landweber algorithm. The simulation results indicate that calculating the electrical conductivity sensitivity requires the extraction of the electric field in the x and y axes, while calculating the magnetic permeability sensitivity matrix requires the extraction of the magnetic field in the z-axis. Meanwhile, it can determine the electrical parameter distribution of the conductors by measuring boundary voltages, but the intensity of the magnetic permeability sensitivity matrix in the object field is significantly higher than that of the electrical conductivity sensitivity matrix, which indicates Dual-Plane Linear Array Electromagnetic Tomography is suitable for capturing boundary magnetic fields to reconstruct images. Furthermore, by comparing the image reconstruction of vehicles in six different positions using the three algorithms, it is found that using the Landweber algorithm provides the best results in reconstructing the magnetic permeability distribution images of vehicles, with the minimum Image Error (IE) of 0.905 and the maximum Correlation Coefficient (CC) of 0.547. Therefore, Dual-Plane Linear Array Electromagnetic Tomography is feasible in simulation and can effectively be used for the magnetic permeability distribution image reconstruction of road vehicles, possessing potential application value in improving road traffic safety management and reducing traffic congestion.