Abstract:Magnetic materials are important stimulus-responsive materials that can be controlled wirelessly by an external magnetic field through tissues and organs. They are widely used in biomedical fields such as medical robots, artificial organs, biochemical synthesis, and drug delivery due to their high biocompatibility, simple magnetic field control, and fast regulation speed. Complex work scenarios and multifunctional requirements pose higher demands for the precise control of magnetic materials, ranging from simple planar driving based on magnetic force to complex spatial driving based on both magnetic force and torque and from functional implementation based on the material′s own motion deformation to achieving more complex environmental detection as a multifunctional flexible electronic device carrier. This article introduces the magnetic properties, magnetic field control platform, and magnetic pole programming techniques of several commonly used magnetic materials, demonstrating the application progress and numerous challenges faced by three different types of magnetic materials: Magnetic liquids, magnetic solids, and magnetic membranes, in the biomedical field. Finally, future development trends are discussed.

Abstract:To solve the problem that power transformer vibration signals is difficult to predict because of the non-stationary characteristic, an autoregressive integrated moving average prediction model based on improved dung beetle optimizer algorithm is proposed. Firstly, ADF test and KPSS test are used to check the stationary of the transformer original vibration signal, and if it is not stationary, differential processing is performed until the signal is stationary. Secondly, the periodic mutation mechanism is introduced into dung beetle optimizer algorithm to improve the optimization ability of the algorithm, and the parameters p and q of autoregressive integrated moving average model are determined by improved dung beetle optimizer algorithm to realize the prediction of transformer vibration signal. Finally, the validity of the proposed model is verified by using the actual collected vibration data of a 0. 4- / 0. 4-kV, 15-kVA three-phase doublewinding dry-type transformer. The simulation result shows that the mean absolute percentage error of the model can reach 3. 77%, while the mean absolute percentage error of the autoregressive integrated moving average model, long short-term memory network, recurrent neural network and convolutional neural network are 5. 34%, 4. 74%, 5. 03% and 5. 40%, respectively. Therefore, the proposed model can achieve accurate prediction of transformer vibration signal.

Abstract:Aiming at the problem that traditional visible light is difficult to realize personnel behavior detection and identity recognition in dark environment, this paper combined with infrared thermal imaging technology to study an algorithm for personnel behavior detection and identity recognition in dark environment based on Baidu Paddle deep learning framework. First, after field collection, the behavioral dataset of infrared thermal imaging personnel totaled 10 900 pieces of 9 behavior categories and the double-light face dataset totaled 3 000 pieces of 30 personnel. In terms of behavior detection, the lightweight network PP-LCNet is used to improve the YOLOv5 backbone network for personnel behavior detection, reducing model parameters greatly and improving detection accuracy and reasoning speed. In terms of face recognition, CycleGAN algorithm is introduced to improve InsightFace to transform infrared faces into visible faces for identity recognition and improve face recognition accuracy in dark environments. Finally, the cascade of infrared human behavior detection network and face recognition network is realized, and real-time behavior detection and identity recognition can be achieved in the dark environment, which has a good application effect. The experimental results show that compared with the original network model, the parameters of YOLOv5 based on PPLCNet are reduced by 56. 4%, the average precision mAP is increased from 89. 1% to 94. 7%, and the reasoning speed is increased from 68 to 101 fps. Based on CycleGAN algorithm, the recognition accuracy of InsightFace is improved from 84% to 99% in the dark environment of the original network.

Abstract:In view of the problem that depth accuracy is affected by camera parameters when disparity is converted into depth in the current disparity estimation network, and depth accuracy decreases sharply at long distance, a full range depth balanced stereo matching network (FRDBNet) is proposed. Firstly, the depth cost volume is constructed to make the network learn the probability distribution of the full distance depth, and the depth is directly generated by depth regression. Then, the training strategy of disparity and depth loss fusion is used to make the network pay attention to the depth estimation of the long, middle and near three segments distance at the same time. Finally, a disparity optimization module is designed based on the seven neighborhood features corresponding to the original disparity right map to further improve the depth estimation accuracy of the network. Experiments on the DrivingStereo dataset of large real-world driving scenarios show that for the full distance[1,100]m depth estimation, the depth accuracy of FRDBNet at[1,30]m short distance,[30,60]m middle distance and[60,100]m long distance is 10. 38%, 15. 11% and 20. 35% higher than that of ACVNet with superior performance of CVPR2022, respectively, achieving a good balance of depth accuracy.

Abstract:In recent years, the growing Internet of Things (IoT) has generated huge amounts of data, which has put enormous pressure on infrastructures such as the network cloud. One of the obstacles to fog computing is how to allocate computing resources in a way that minimizes network resources. A heuristic-based TCC (time cost computing-power) algorithm is proposed to optimise the task scheduling problem in genetic algorithm-based “cloud-fog” computing in this heterogeneous system, including execution time, operational cost and total computing power resources. The algorithm is based on “ cloud-fog-end-net ” hybrid computing task scheduling, and uses evolutionary genetic algorithms as a research tool, combining the advantages of cloud computing, fog computing and genetic algorithms to achieve a balance between latency, cost and computing power. In the hybrid computing task scheduling, this algorithm has a better balance performance than TCaS algorithm which only considers a single metric; the adaptation value of this algorithm is 0. 93% and 26. 02% higher than BLA algorithm and RR algorithm respectively. The algorithm is also flexible enough to match the user’ s needs in terms of high performance-cost-computing power, enhancing the effectiveness of the system.

Abstract:Accurate and efficient semantic description of human posture is integral to human behavior recognition. It is also a key to quickly understanding individual states and events. In recent years, human key point detection technology has gained significant development. However, the research on the semantic description of the human pose has not attracted enough attention. To this end, we propose a geometric statistics-based semantic description method for human posture. Firstly, the obtained human key points are divided into several sets. Then, the geometric distribution characteristics of each key point set are calculated to describe the human posture. Finally, the semantics of the human pose is judged using a hierarchical strategy. This method employs the idea of the set to improve the robustness of recognizing human posture. The experimental results on multiple real scene datasets show that the proposed method attains an average accuracy of 90. 8% and 77. 1% for identifying human pose on the IFD and PASCAL datasets for simple and complex singleperson pose, respectively, and 77. 2% on the MPII dataset for the complicated multi-person pose, which are better than the performance of compared approaches. In conclusion, the proposed method can achieve more accurate human pose semantic descriptions despite the absence of some key points.

Abstract:In order to achieve an intelligent prosthetic hand that can naturally simulate the continuous motion of a human hand, this paper proposes a DF-ANN model based on sEMG to estimate the finger joint angle. The method introduces the SE-Net module in the channel attention mechanism to enhance the relevant feature expression of sEMG, reduce the loss of essential features of sEMG, and effectively improve the performance of the regression model. 10 healthy subjects were selected for experiments with 10 different hand gestures, and regression measures such as R-Squared (R 2 ) were chosen to evaluate the accuracy of the method’s joint angle estimation. The experimental results showed an R 2 of 86. 5%. Compared with the DF-ANN model without introducing SE-Net, the deep forest, and an artificial neural network alone, the R 2 is improved by about 4%. It indicates that the method effectively reduces the error of successive decoding of joint angles of sEMG and can contribute to the supple control of intelligent prosthetic hands.

Abstract:Stochastic configuration networks (SCNs) have been successfully applied to big data analysis with their general approximation capability and fast modeling properties. Based on the SCNs, stochastic configuration networks with block increments (BSC) use block increment mechanism to improve the training speed, but increase the complexity of model structure. To solve the above challenges, an adaptive switching learning model based on forgetting factor stochastic configuration networks (FSCN-I and FSCN-II) with (ASLM) is proposed. FSCN-I adjusts the size of node blocks by error values and forgetting factors to improve the training speed, and FSCN-II introduces a node removal mechanism to reduce the complexity of the model structure. ASLM consists of FSCN-I and FSCN-II, both of which are randomly switched according to the adaptively changing boundaries to improve the training speed of the model and the complexity of the model structure is reduced based on FSCN-I. Finally, the effectiveness of the method is demonstrated with the underlying dataset and industrial examples.

Abstract:Aiming at internal leakage failures of gas transmission pipeline valves in petrochemical industry, this paper proposes an identification method of acoustic emission signal of valve internal leakage based on full convolutional neural network ( FCN) by combining acoustic emission detection technology with deep learning technology. The method uses acoustic emission technology to collect acoustic emission signal of valve internal leakage and builds a valve internal leakage classification and diagnosis model based on FCN, which fully exploits the superiority of acoustic emission technology in the field of valve internal leakage detection and the high performance of FCN in time series classification tasks. Compared with traditional identification methods, this method does not require any feature extraction or complex preprocessing of the original collected data. Instead, the task of feature extraction is also handed over to the neural network model to learn and complete, which can realize end-to-end classification and identification of valve internal leakage acoustic emission signal. The data sets of valve internal leakage acoustic emission signal are collected and produced through the experimental platform of valve internal leakage detection, and the binary classification model of valve internal leakage acoustic emission signal based on FCN is established as well. The experimental results show that the accuracy of classification and recognition of the model can reach 98. 72%. Compared with other advanced classification models, the model shows more superior recognition performance on the data sets and has higher training efficiency, which also has good anti-interference performance against environmental noise at the same time.

Abstract:To address the problem that Lamb wave damage probability imaging requires a baseline signal, a baseline-free damage probability imaging algorithm for Lamb waves based on the same propagation distance is proposed. Firstly, the power spectral density of the signal on each path is solved, then the power spectral density values on the paths with the same propagation distance are grouped, the path corresponding to the maximum value of the power spectral density in each group is regarded as a non-damaging path, and compared with other power spectral density values in the same group to construct a damage index, and finally the damage probability imaging algorithm is combined to identify the damage location. The results show that this baseline-free damage imaging algorithm is capable to accurately localize defects within 6% relative error for localizing multiple types of damage within the inspection area.

Abstract:Vortex flowmeter is a fluid vibration flow meter, which is easily affected by pipeline vibration and flow field disturbance. The measurement error is large under the condition of small flow rate, especially under the interference of strong vibration, vortex signal is submerged, and it cannot be accurately measured. In this paper, a vortex flowmeter system which is resistant to strong periodic vibration is developed. Differential charge amplifier with negative voltage feedback is used to improve the ability of discharge flow signals. A frequency variance algorithm based on frequency shift strategy is proposed to reduce the influence of strong periodic vibration. Firstly, the influence of similar frequencies is reduced by frequency shift strategy. Then, according to the frequency band width of traffic signal and periodic vibration interference, the frequency of traffic signal is determined by calculating and comparing the frequency variance. The signal processing system of vortex flowmeter is developed and tested. The results show that the developed system extends the lower limit of range and can accurately extract signals under the condition of strong periodic vibration interference, and the accuracy is improved by one order of magnitude.

Abstract:Since the wind speed and direction measurement accuracy of ultrasonic anemometers is low and the noise suppression capacity is poor, a double-array receiving array ultrasonic sensor wind speed and direction measurement technique based on the secondary correlation method is presented. Firstly, an array structure consisting of an ultrasonic transmitting array and two receiving arrays is adopted, secondly, a method for measuring ultrasonic propagation time based on quadratic correlation is given according to the structure of the system, which can effectively improve the accuracy of wind speed and direction measurement by using stronger performance of the quadratic correlation algorithm for noise suppression. Finally, the effectiveness of the proposed method is verified by simulation experiments and a real measurement system built. The experimental results show that the wind speed and direction measurement method of the dual-array receiving array ultrasonic sensor based on the quadratic correlation algorithm has a strong noise suppression capability, and has a higher wind vector measurement accuracy. The practical test shows that the maximum error of wind speed and wind direction angle measurement are 0. 24 m/ s and 2. 4°, which basically meets the accuracy requirements of ultrasonic wind speed and direction measurement.

Abstract:Conducting a quantize value of driving risk is crucial for the human-like driving decision of intelligent vehicles (IV). Aiming at the challenge of quantifying driving risks in complex multi-task scenarios, a method of driving risk quantification of IV based on human risk perceived mechanism is proposed. By utilizing vehicle or road sensors, measurements of the surrounding environment and state information have been obtained. And a cost map of the driving scene is created by assigning costs to potential collision factors such as roads, plants, and obstacles that the driver believes may occur in the first stage. Based on the fundamental principles of human driving and vehicle motion states, a dynamic risk model is established utilizing Gaussian functions. The real-time computation of driving risk with human-like characteristics is accomplished through the integration of the cost map for the driving environment and a dynamic risk model. The simulation results demonstrate the effectiveness of the proposed method in quantifying dynamic driving risk based on human driving experience, which is applicable to autonomous driving decision-making for IV and capable of generating human-like driving behavior.

Abstract:In the field of vacuum leak detection in fusion devices, the future fusion devices are operated with tritium and the leak checkers do not have access to the devices for leak checking, which makes this task extremely difficult and time-consuming. In order to realize the fast and accurate detection of fusion device leakage equipment,and realize the fast and accurate detection of fusion device leakage equipment, this paper takes the six-degree-of-freedom robotic arm as the research object, and proposes a GV2-YOLOv5 vacuum equipment detection method for vacuum leakage detection robots to identify and locate the vacuum equipment for helium injection. In this method, the C3GhostV2 module is constructed by combining lightweight GhostNetV2 network, while using lightweight GhostConv to extract target features, thus reducing the number of model parameters and improving the computational speed. Bottleneck Transformers and ECA Attention mechanism are added to the feature fusion network to improve the network feature extraction capability and to enhance the model channel features. The experimental results show that the average accuracy of the improved model is 93. 2% on the homemade dataset, which is 1. 4% higher than YOLOv5s, the amount of model parameters is reduced by 29. 5%, and the detection speed is 92 fps, which meets the requirements of real-time and accuracy, and provides a solution for the vision localization technology of vacuum leak detection robot.

Abstract:In order to realize quantitative imaging of pipe wall corrosion defects, an imaging algorithm based on physics embedded convolution neural network is proposed to reconstruct pipe wall thickness from ultrasonic guided wave signals. Firstly, the twodimensional acoustic wave model of ultrasonic guided wave propagation on the pipe wall is derived. The wave equation in frequency domain can be solved by matrix LU decomposition to realize the forward modeling from the pipe wall guided wave velocity diagram to the acoustic field signal. Secondly, the physics embedded convolution neural network is built, including three iterative layers, each of which is composed of forward model and residual inversion subnetwork. The pipeline simulation data set containing random corrosion defects is generated, and the network is built for training and inversion. The average Pearson correlation coefficients of the imaging results of the training set, verification set and test set are 94. 91%, 86. 47% and 87. 37% respectively, and the defect image consistency is high. The experimental system is built, and the guided wave signal is collected on the pipe with irregular step defects for inversion. The imaging results is remarkable, with a mean square error of 0. 005 7 for the thickness map. The algorithm combines the physical model with neural network to achieve high-precision imaging from guided wave signal to pipeline thickness map.

Abstract:Because of the inherent structural defects of P3P problem, the minimum solution method for calibration of 2D Lidar and camera has some shortcoming, such as poor numerical stability and accuracy. Aiming at the above problem, this paper proposes a robust minimum solution calibration method, which improves algorithm of P3P problem and the error measurement of optimal solution selection. Firstly, according to the P3P problem constructed by three checkerboards, the enhanced RP3P algorithm is used which can improve the stability of the solution. Secondly, an optimal solution selection strategy based on observation probability with uncertainty weighted is designed to improve the accuracy of the optimal solution. Experimental results show that the proposed algorithm can significantly improve the probability of reasonable solution and calibration accuracy compared with the algorithms in the literature. Under different noise levels, compared with Francisco method and Hu method, the probability of reasonable solution is improved by 5% ~ 41% and 2% ~ 20%, the rotation matrix accuracy is improved by 2° ~ 6° and 1. 5° ~ 2°, the translation vector accuracy is improved by 180 ~ 520 mm and 150~ 180 mm, so the performance is improved obviously.

Abstract:Reliable localization is a crucial prerequisite for robots to perform navigation and path planning. Traditionally, locations of robots are derived from the ranging measurements between ultra-wideband (UWB) tag and anchors, results with poor accuracy may be yielded when available anchors are insufficient. To tackle this issue, a mobile robot localization method based on ultra-wideband bearing and range is proposed. Firstly, the direction of the UWB tag, i. e. , the forward field of view ( FOV) or behind non-field of view (NFOV) of the anchor, is determined according to the standard deviation of the UWB bearing signal, thus eliminating the front-back singularity in the robot localization process. In addition, constraint functions are constructed utilizing UWB range and bearing measurements, and global pose optimization is achieved by fusing odometry and UWB measurements through a graph-based optimization algorithm. The experiment results show that the proposed method has strong robustness and is able to locate the irregularly moving robot with a localization accuracy of 0. 093 m in a 13 m×6 m indoor environment, which is 46% better than the traditional localization method based on ranging UWB and odometry fusion.

Abstract:The defect of the PDC bit compact is an important factor affecting the drilling efficiency, and detecting whether the PDC bit compact is defective is a prerequisite for repairing the PDC bit. In order to reduce the false detection of PDC drill bit composites and improve the detection accuracy, a target detection algorithm based on improved YOLOv5 is proposed. This method is based on the YOLOv5 network, and integrates the RepVGG reparameterization module to enhance the feature extraction ability of the network; introduces the coordinate attention mechanism in the C3 module, embeds the position information in the channel attention mechanism, and improves the target detection ability of the defective composite film. Improve the bounding box regression loss function to the WIoU loss function, and formulate a suitable gradient gain allocation strategy. The experimental results show that the precision rate of the improved network increased with 2%, the recall rate increased with 0. 9%, and the mean average precision ( mAP) increased with 1. 3%, reaching 98%, which can realize the defect recognition of PDC drill bit composites.

Abstract:Temperature changes have a great impact on guided wave propagation, but the composite error caused by temperature and other factors on guided waves is rarely studied, and pipelines are often in various severe environments. It is necessary to study the composite error between temperature changes and different working conditions. Regarding the relationship between the water-filled pipeline and the temperature change, a new compensation idea is proposed, that is, the signal set is used to match the similar baseline signal, and the water filling information is judged at the same time, then the time domain part of the monitoring signal is stretched by the baseline stretching method, and finally the corresponding maximum residual amplitude and the water-filling deviation are subtracted to complete the composite error compensation of the signal. The feasibility of the method was analyzed by COMSOL simulation, and experiments were designed to verify the effect. The experimental results show that the average amplitude of the residual signal after compensation is 5 dB lower than the average amplitude of the residual signal after compensation, that is, the method can effectively compensate for the composite error caused by temperature change and water filling.

Abstract:In order to solve the problems in the current infrared and visible image fusion algorithm, such as missing scene detail information, unclear target region detail information and unnatural fusion image, an attentional residual dense fusion network (ARDFusion) for infrared and visible image fusion is proposed. The whole architecture of this paper is an auto-encoder network. First, the encoder with the largest pooling layer is used to extract multi-scale features of the source image, then the attention residual dense fusion network is used to fuse the feature maps of multiple scales. The residual dense blocks in the network can continuously store features and maximize the retention of feature information at each layer. The attention mechanism can highlight target information and obtain more detailed information related to the target and scene. Finally, the fused features are input into the decoder and reconstructed through upsampling and convolutional layers to obtain the fused image. This article proposes an attention residual dense fusion network for infrared and visible image fusion. The experimental results show that compared to other typical fusion algorithms in existing literature, it has better fusion performance, can better preserve the spectral characteristics in visible images, and has significant infrared targets. It has achieved good fusion performance in both subjective and objective evaluations.

Abstract:Numerical simulations were conducted to investigate three non-standard design or non-ideal situations that may arise in the design, fabrication, and usage of PPD laminar flow elements. For the non-standard design with different numbers of capillaries in two branches, in which the numerical simulations were conducted, the pressure drop curves and differential pressure-flow relationships obtained showed minimal differences compared to the standard structure. The relative deviation between the calculated flow rate and the accurate flow rate, as determined by the pressure drop calculations, was within ±0. 4%. Additionally, the experimental data from flow rate measurements confirmed the simulation results, further demonstrating the applicability of the PPD principle to structural designs with varying numbers of capillaries in the two channels. Regarding the potential occurrence of blockages at the capillary inlets or outlets during fabrication or usage, calculations revealed the presence of nonlinear deviations in the differential pressure. Depending on the location of the blockage, either positive or negative deviations may occur. Furthermore, when there is a deviation in the capillary diameter between the two branches, the impedance characteristics of the two branches become dissimilar. This dissimilarity in flow rates and velocities within the capillaries of the two branches leads to incomplete compensation of local resistances at the inlets and outlets, causing the differential pressure-flow curve to deviate from the ideal curve. The aforementioned research findings hold valuable insights for the practical application of PPD laminar flow sensing technology.

Abstract:In the context of wind turbine gearbox fault early warning, this paper proposes a method based on graph attention and temporal convolutional networks to address the issue of insufficient data information mining. By establishing physical connections for each feature point in both temporal and spatial scales, the method expands the feature dimension to enhance fault warning accuracy. Graph attention network captures spatial relationships, while temporal convolutional network improves temporal feature capturing. Experimental results using real data from a wind farm show that the proposed method can issue fault warnings 122 hours in advance, outperforming other methods by 52 to 63 hours with reduced prediction errors (1. 05% to 3. 76%). The approach also enhances result interpretability using t-SNE and probability density curve analysis.

Abstract:As a key component of the main power output of harvester, the fault performance of the spiral bevel gear is usually the excitation impulse. To monitor and diagnose the faults of the main transmission gearbox of agricultural harvester timely and effectively, an improved wavelet packet decomposition ( WPD) method based on dynamic model driven optimization is proposed in this paper. Aiming at the multi-component modulation phenomenon of gear damage and the characteristics of wavelet basis function specific timefrequency window to analyze the signal, the proposed method establishes the physical model of gear dynamic damage to assist in screening of wavelet packet decomposition coefficients that adapt to the gear damage characteristics. Thus, the wavelet basis function selected for the decomposed signal is optimized, so that it has a better ability to extract feature information of gear fault. Through the envelope spectrum analysis of the experimental signal and fault signal of the Chinese onion harvester gear, it is verified that the proposed method can be effectively applied to the fault diagnosis of the harvester gear.

Abstract:Non-metallic pipelines are widely used in the construction of underground pipe networks. As one of the main detection methods of underground public facilities, the ground penetrating radar method is often used in the detection of underground non-metallic pipelines. Due to the weak echo signal of non-metallic pipelines, complex underground structures, and difficulty in accurately estimating wave velocity in the detection process, radar detection imaging and positioning have become a problem. In response to the above questions, the wave velocity estimation and echo signal processing methods suitable for buried non-metallic pipelines are studied, the spatial frequency domain interpolation imaging and binarized positioning algorithms are proposed, the feasibility of the imaging positioning method is verified by numerical simulation, and the field tests of non-metallic pipelines buried at 0. 400, 0. 600 and 1. 100 m are carried out. The experimental results show that the proposed method can effectively eliminate the hyperbolic effect, and the imaging positioning errors in the detection area are 0. 006, 0. 006 and 0. 012 m, respectively, which meet the positioning requirements of buried non-metallic pipelines. The research results can provide technical support for the imaging and positioning of buried non-metallic pipelines and are conducive to improving the detection accuracy of ground penetrating radar on underground non-metallic pipelines.

Abstract:A motion planning method based on the angular velocity and linear velocity generated by the convex optimized smoothing dynamic window approach is proposed for the problem that the angular velocity and linear velocity sequences generated by the traditional dynamic window approach fluctuate frequently and cannot control the normal motion of the robot due to the rugged road surface under the coal mine. Firstly, the dynamic window method is used to generate a set of angular velocity linear velocity sequences, then the objective function is constructed based on this sequence to obtain the optimal solution using convex optimization to remove the noise in the original sequence and achieve the motion control of the robot, and the obtained optimal solution is the smoothed angular velocity linear velocity sequence. Finally, simulations are performed using matlab to verify the smoothing effect of convex optimization under different maps and to compare with mean filtering and VMD reconstruction of the signal. The results show that, compared with the comparison algorithms, the convex optimized dynamic window method has the best smoothing effect on the angular and linear velocity sequences with the smallest average absolute errors of 0. 005 6 and 0. 001 8, smoothing the signal while preserving its characteristics.

Abstract:In monocular 3D detection, the complex network structure and inaccurate target depth information obtained after depth estimation are two problems that need to be dealt with. To address this issue, we propose an end-to-end joint multi-attention depth estimation monocular 3D target detection network structure, named CDCN-3D. First of all, to obtain the salient features of the target, we introduce an adaptive spatial attention mechanism to aggregate the pixel features, which enhances local features and improves the network representation ability. Second, we use an improved C-ASPP approach to address the problem of local information loss in depth estimation, capturing more accurate direction perception and position-sensitive information for each depth information. Finally, the accurate P-BEV is used to map the three-dimensional information of the target to a two-dimensional plane, and then the single-stage target detector is used to complete the detection and output task. Through experiments on the KITTI dataset, the proposed CDCN-3D network shows improved accuracy compared to other networks, with the same FPS as that of the existing monocular 3D detection network. More specifically, and the detection accuracy of the CDCN-3D network is improved by 2. 31%, 1. 48%, 1. 14% respectively by the class of Car、Pedestrian、Cyclist, which can complete the 3D target detection task.

Abstract:Creepage distance of rod composite insulator is an important concern in its quality control. A new method based on machine vision for measuring creepage distance of composite rod insulator is proposed in this paper. Firstly, the image acquisition platform of rod composite insulator is designed, and its axial continuous images are obtained. The SIFT algorithm is used to splice the obtained continuous images, and the Canny edge detection algorithm is used to build the recognition model to realize the automatic measurement of the creepage distance of the rod suspension composite insulator. The test shows that the standard deviation of repeated measurement of image measurement method is about 5% ~ 12% of that of traditional method, and its measurement efficiency is more than 6 times higher than that of traditional method. The accurate and efficient measurement of creepage distance of rod composite insulator is realized.