Abstract:Motion recognition is one of the key links in the intelligent monitoring of ski jumping. This paper takes ski jumping as the research object, and fuses the data of different inertial sensors and different joint nodes to generate structured data by stacking, to realize the recognition of ski jumping movements by using deep convolutional neural network. Firstly, the collected inertial sensing data of different sensors and different human body points in the process of ski jumping are normalized and mapped to between ［0,1］. And then using color mapping to stack all kinds of data to create an image. Then use two-dimensional convolutional neural networks such as Resnet, to identify 5 types of movements in ski jumping: start-to-slip, straight-line assist, curve-assist, take-off and early flight, stable flight and landing. The experimental results show that the 2 250 stacked inertia signal images generated by 9 times of ski jumping data fusion are recognized, and the recall rate and accuracy are 93.8% and 91.7%, respectively. At the same time, the influence of a single class inertial sensor on the recognition result of the fusion data of each joint node is analyzed. The proposed method of stacking inertial signal fusion and action recognition of different sensors and different joints can provide support for intelligent analysis of ski jumping.

Abstract:Gesture recognition is the key to human-computer interaction. In order to better realize the fusion of EEG and EMG signals and accurately recognize human motion, this paper established a research system of real-time gesture detection and recognition based on Grael EEG amplifier. The surface electromyography (sEMG) signals of eight different gestures from five channels were collected by Grael EEG amplifier and Curry8 system, and the collected sEMG signals were preprocessed by filtering and denoising, sliding window segmentation and feature extraction. Finally, several commonly used classifiers and convolutional neural network (CNN) are used to classify and recognize sEMG signals of different gestures in real time. The results show that the recognition accuracy of CNN is the highest, reaching 92.98%. After 30 times of real-time recognition and detection for each gesture, the results show that the recognition delay is about 1~1.5 s, and the accuracy of real-time recognition can be up to 90%. This system provides a feasible method to study the fusion of EEG and EMG signals in the future, and shows great potential and application space in human-computer interaction.

Abstract:Techniques based on deep learning for hyperspectral image classification can effectively extract features and promote the mining and utilization of the rich information. The performance of existing methods is still limited by the insufficient extraction of shadow information and the inefficient use of features. For information extraction in shadow areas, dynamic stochastic resonance can enhance the signal by using noise to improve the ability of information expression. For feature utilization, the attention mechanism is embedded in the convolutional neural network, which can further extract and fuse from the space dimension and channel dimension based on the high-level features extracted, screen out features more critical to the current task target, so as improving the classification performance. The experimental results show that dynamic stochastic resonance can effectively enhance signal, the classification accuracy on real world dataset Hydice is improved from 96.48% to 97.14%, and is improved by 0.408 4% with convolutional attention block added. Further verification by comparison with other methods, the classification accuracy on Hydice, Indian Pines and Pavia University reaches 97.436 1%, 99.219 5% and 99.929 9% respectively, which has obvious advantages. It is proved that the method is effective and has good classification performance, and has broad application prospects in the field of hyperspectral image classification.

Abstract:Unmanned aerial vehicle photogrammetry without ground control points can effectively improve production efficiency and reduce production costs. This method shows great advantages in areas inaccessible to people. However, there are a number of problems with this method. One of the main problems is that the accurate camera parameters cannot be obtained through the on-the-job calibration method. Inaccurate camera principal distance can seriously affect the object point elevation. The other one is that the projection deformation and earth curvature also have impacts on the elevation accuracy, when the mapping task is carried out in the map projection frame. Therefore, this article analyzes the specific causes of elevation errors and realizes the camera self-calibration in a geocentric frame with control strips. Then, the exterior orientation elements of the images are calculated in the map projection frame without control strips. Finally, the elevation errors caused by map projection deformation and earth curvature are corrected. The experimental results show that the elevation RMSE of the two data sets decreased from 0.298 m and 0.374 m to 0.075 m and 0.080 m, respectively. Elevation accuracy are less than 0.1 m. Therefore, the method in this paper can achieve accurate mapping in a map projection coordinate system without ground control points.

Abstract:Aiming at the problem of high difficulty and low accuracy in road traffic sign recognition under haze weather, a traffic sign recognition model based on YOLOv5 was proposed. Firstly, the convolutional attention mechanism was integrated into the original YOLOv5 model to enhance features in the spatial dimension and channel dimension to suppress the interference of haze weather on the model. Then, BiFPN is used as the feature fusion structure in neck layer to more fully fuse multi-scale features and reduce the loss of target information. CIoU is used as the loss function of YOLOv5 to improve the positioning ability. K-means clustering algorithm was used to re-obtain anchor frame values in TT100K and CODA datasets to accelerate the convergence speed of the model. The experimental results show that the recognition accuracy of the improved model reaches 92.5%, which is 5.6% higher than that of YOLOv5, and it can still accurately identify traffic signs in haze weather. and the speed can reach 27 FPS, which can be used for real-time detection.

Abstract:In view of the lack of feature extraction ability and low recognition efficiency of common convolutional neural network in facial expression recognition, this paper proposes a facial expression recognition based on multi-scale feature fusion of MobileNetV3. Firstly, MobileNetV3 was used for feature extraction to obtain high-level emotion information. Secondly, the DenseNet structure is used in the backbone network to enhance feature reuse and improve the expression ability of important facial features. Then the feature pyramid module is used to fully obtain the deep and shallow multi-scale fusion features of face images, so as to improve the feature extraction ability and real-time performance of MobileNetV3. Finally, the full connection layer is used to construct a classifier to classify the facial expression, so as to complete the facial expression recognition. Through experimental verification, the results show that the recognition accuracy on CK+and FERPlus datasets can reach 88.3% and 98.8%, which are improved by 2.3% and 1.5% respectively compared with the existing methods, indicating that the proposed method has good recognition effect and strong generalization.

Abstract:Aiming at the shortcomings of the polynomial regression color correction method, an adaptive weighted root polynomial regression algorithm is proposed. In the process of polynomial regression color correction, it is necessary to manually calibrate the position of the color block of the color card, which is complicated and prone to human error. In view of the problem that the high-order term of the polynomial will amplify the noise and is not robust to the noise, the algorithm in this paper will adaptively adjust the weight matrix to reduce the influence of singular values on the fitting performance, and then calculate another gain coefficient matrix from the color difference value, thereby improving the correction accuracy. It has been verified by experiments that the algorithm in this paper has a great improvement in the CIELab color difference value and PSNR compared with the traditional polynomial regression method. Among them, the average CIELab chromatic aberration value of the traditional polynomial regression method is as high as 6.5, which is greatly affected by the environment. However, the chromatic aberration value of the proposed algorithm can be stabilized below 3.2 after correcting images in different environments.

Abstract:Aiming at the action recognition of subtle actions in the similar background of drivers, X3D-M-GC-AE based on X3D network is proposed. By introducing the lightweight self-attention network GCnet, the attention to key features in time and space is improved, and the detection accuracy is improved without increasing parameter quantities. Action enhancement block is designed to make the network more sensitive to the action information in time series. Introducing knowledge distillation, taking X3D-XL as the teacher network and X3D-M-GC-AE as the student network, so that X3D-M-GC-AE can be used in real vehicles with less parameters and calculations. The experimental results show that the maximum test accuracy of teacher network can reach 75.56%, and that of student network can reach 71.13%. This framework can achieve high-precision detection results in the case of low requirements for vehicle hardware equipment.

Abstract:In view of the shortcomings that the 2S-AGCN model of the two-stream adaptive graph convolutional network ignores the long-distance information of features and channel dependence in human motion recognition, a dual attention mechanism is designed to improve the graph convolution module of the 2S-AGCN model to improve the accuracy. The dual attention mechanism includes the spatial attention mechanism and the channel attention mechanism. The spatial attention mechanism selectively focuses on the context. The channel attention mechanism is divided into two parallel modules. The first part improves the distinguishability of features. The second part preserves accurate location information while capturing the remote dependency of features. The results show that the model based on the two-stream adaptive graph convolutional networks 2S-AGCN, which incorporates the dual attention mechanism module, has improved Top-1 and Top-5 on the Kinetics dataset by 0.6 and 1.3 percentage points respectively, Top-1 on the CS and CV of NTURGB+D120 dataset by 1.2 and 0.5 percentage points respectively, and Top-1 on the CS and CV of NTURGB+D dataset by 0.2 and 0.1 percentage points respectively.

Abstract:In order to solve the current assembly robot vision system’s problems of high false detection rate, low efficiency, and difficulty in obtaining effective positioning information. a component vision recognition and positioning method based on deep learning was proposed. Firstly, a high-precision detection algorithm based on deep aggregation and decoupling head was designed to improve the accuracy of component identification and subject detection. Secondly, the rules of labeling and determination were designed, and the position subject outlines and grasping points were refined. Lastly, a lightweight detection algorithm based on network pruning was designed to accomplish model compression and improve the efficiency of pin detection and assembly point positioning. The research results show that the method has achieved better performance in the identification and positioning of components. The average error rate of category recognition is merely 0.27%. The calculation is reduced by 29.8%, and the volume of parameters decreased by 22.7%. Through this method, traditional component contour detection is extended to grasp point and assembly point positioning to obtain abundant category and position guideline information, laying a foundation for industrial robots to grasp and assemble accurately, reliably and stably.

Abstract:Aiming at the problems of unstable convergence and poor optimization effect in the multi-objective task scheduling of deep reinforcement learning in the edge computing environment, a new multi-objective task scheduling algorithm based on an improved competitive deep double-Q network (IMTS-D3QN) was proposed. First, the selection and calculation of the target Q value are decoupled by the deep double-Q network to eliminate overestimation, the immediate reward experience sample classification method is adopted to extract experience samples from the experience replay unit, which improves the utilization rate of actual samples, which speeds up the training speed of the neural network. Then, the neural network is optimized by introducing competing network structures. Finally, the soft update method is used to improve the stability of the algorithm, and the dynamic ε-greedy exponential decreasing method is used to find the optimal strategy. The Pareto optimal solution is obtained through different linear weighting combinations to minimize the response time and energy consumption. The experimental results show that, compared with other algorithms, the IMTS-D3QN algorithm has obvious optimization effect in response time and energy consumption under different number of tasks.

Abstract:As the number of motor vehicles continues to increase the problem of traffic congestion in cities becomes more and more apparent, an intelligent traffic system is designed in this paper to alleviate traffic congestion. In the vehicle detection section: The problem of ghosting from ViBe background modelling is solved using background mean modelling, and the adaptive background update strategy is implemented by assigning different background update rates to backgrounds of different complexity. Traffic light timing section: To address the problem of single-stage fuzzy control with queue length as the fuzzy control input error, a two-stage fuzzy controller based on congestion intensity is constructed, and the timing scheme is derived after fuzzy inference and clarification, so as to make adjustments to the green light time. The experimental results show that in the traffic flow detection part: Through the testing of different types of traffic scenarios, the comprehensive accuracy of the improved ViBe algorithm in vehicle flow is improved by 11% compared to the ViBe algorithm, which can provide accurate data support for the timing strategy. In the traffic light timing part: Compared with the existing traffic light timing methods, the signal light timing strategy based on two-level fuzzy control proposed in this paper reduces the average vehicle delay time and the average vehicle travel time by more than 3.34 and 5.65 s respectively under three traffic scenarios, which can play a role in easing traffic congestion.

Abstract:In order to ensure that the intelligent vehicle can run safely according to the planned path and meet the vehicle dynamics characteristics, and solve the unnecessary search problems caused by the lack of guidance strategy in the hybrid A* algorithm, a probabilistic A* algorithm is proposed to obtain the search rough path firstly, which improves the search efficiency in the subsequent search process. Secondly the path points obtained by the probabilistic A* algorithm are used to guide the search direction to avoid colliding with obstacles. Finally, optimize the cost function of nodes. Simulation results show that, compared with the hybrid A* algorithm, the proposed algorithm reduces the search time by 10.8% on average, and the resulting path is relatively regular and smooth. The algorithm can plan a safe, feasible and smooth path for intelligent vehicles in a short time.

Abstract:Aiming at the shortcomings of the traditional A* algorithm in UAV path planning, such as low efficiency, a large number of redundant path points, and many path transitions, an improved A* algorithm based on two-way mechanism is proposed. First, a bidirectional search mechanism is introduced, which takes the starting point of the opposite search as the end point, and then determines that the end point is located in the quadrant of the starting point for bidirectional search, so as to improve the search efficiency. Finally, the path smoothing strategy is introduced to smooth the initial path obtained by the bi directional search to reduce redundant path points and turning points. The traditional A* algorithm and the improved A* algorithm are compared through the MATLAB platform. The experimental results show that compared with the traditional A* algorithm, the improved A* algorithm proposed reduces the average path planning time by 61.61%, path points by 83.09%, and path turning points by 46.97%, which can effectively improve the working efficiency of UAVs and generate smooth paths.

Abstract:Insulator is one of the important components of overhead lines. When there is a fault, it will affect the safe operation of power grid. In order to realize rapid and accurate identification of insulator fault, an insulator fault detection method based on improved YOLOv3-Tiny is proposed. Firstly, in order to enhance the small target detection ability, the shallow feature map and the feature map before the second detection layer are spliced in the same dimension to construct the third prediction layer. Then, the network uses Ghost module to replace the convolution layer in the backbone network and reduce the parameters of the model. Then, a new attention module MECA (multiscale efficient channel attention) is designed to enable the network to focus on the salient characteristics of insulators. Finally, a new effective intersection over union (EIoU) is proposed as the frame regression loss function to better locate the insulator position. The experimental results show that the average accuracy (MAP) of the improved YOLOv3-Tiny algorithm in insulator fault detection is as high as 96.1%, which is 17% higher than that of the original YOLOv3-Tiny.

Abstract:Aiming at the problems of slow convergence speed, low search efficiency and random sampling of twoway fast expanding random tree (RRT-CONNECT) algorithm in complex environment with multiple obstacles, this paper proposes an RRT-CONNECT algorithm based on ellipsoid subset sampling. Firstly, on the basis of traditional RRT-Connect algorithm, combined with target paranoid sampling strategy and the advantage of sampling ellipsoid subset, construct a new sampling method, sampling area for constraint, on this basis to find the optimal path from the starting point to the target point point set, and the path as the initial path, by introducing the path pruning algorithm based on triangle inequality, to continuously optimize paths in the iterative process. A path with low cost and no collision was obtained from the starting point to the target point. Finally, a smooth path with continuous curvature was generated by combining the path optimization with the quintic polynomial difference algorithm, so that the manipulator could reach the target point quickly, accurately and stably along the optimal path. Experimental results show that compared with the original RRT-Connect algorithm, the average planning time efficiency is improved by 30.5%, the average sampling points are reduced by 76.74%, and the average path length is shortened by 13.22%. The algorithm has faster convergence speed, higher search efficiency and more significant path optimization effect in the planning process.

Abstract:A method of combining independent frequency sources of the same frequency and phase is proposed to improve the phase noise of the frequency source. Due to the inconsistency of the hardware, however, there is a certain phase difference between independent sources, so the phase noise of the combined frequency source is not close to the theoretical value and is even degraded. This paper proves that combining multiple frequency sources the frequency and phase of which are equal or approximate can improve the phase noise performance of the frequency source, and carries out the theoretical simulation with MATLAB. Aimed at the problem of phase non-synchronization or phase difference between two frequency sources, an adaptive tracking algorithm was proposed to realize the phase synchronization between two different frequency sources. Finally, this paper provided the theoretical simulation results of the proposed algorithm. From the theoretical proof and simulation results, relative to the initial frequency source, the phase noise of the combined frequency source signal is improved by 3 dB. In this paper, an adaptive tracking algorithm was also proposed. The algorithm can control the phase difference between the two initial frequency sources to be less than 10°, the expected effect of which has been achieved.

Abstract:Because of the high complexity of urban traffic flow, the non-repetitive interference in the road network will degrade the dynamic performance of the iterative learning traffic control system. Therefore, a hybrid control method based on Kalman filter and iterative learning is proposed to further improve the robustness and anti-interference ability of the control system. Firstly, the Kalman filter is used to observe the state of the system, and the optimal state of the system is estimated under the condition of random noise. Secondly, an iterative learning control method with forgetting factor is designed, which can enhance the antiinterference ability of large disturbance, and then the reference trajectory of the system is gradually tracked by iterative learning. Finally, the convergence of the algorithm is proved mathematically, and the simulation results also show that the proposed method can reduce the influence of interference on the control system in the disturbance environment, and improve the road capacity and reduce traffic congestion to a certain extent.

Abstract:In order to improve the accuracy of EEG emotion recognition, extract richer feature information and improve the stability of network model, an improved EEG emotion recognition model based on multi-level attention mechanism is proposed. In the aspect of feature extraction, the original EEG signal was transformed into four-dimensional space spectrum time structure to extract rich EEG information. In the aspect of network model, a two-way convolution neural network was constructed to learn spatial and frequency information. It can effectively extract multi-scale features and increase the network width to learn richer feature information. After the convolution layer and pool layer, the batch normalization layer was integrated to prevent over fitting. Finally, a multi-level attention mechanism-bidirectional gated recurrent unit module was constructed to process the time characteristics and cooperate with Softmax classification. The bidirectional gated recurrent unit was used to learn more comprehensive upper and lower level feature information. The multi-level attention mechanism was used to correlate different time slices with the overall time slices in four-dimensional features. The evaluation experiments were carried out in two dimensions of arousal and potency of DEAP data set. The average accuracy of two classifications were 96.38% and 96.73% respectively, and the average accuracy of four classifications was 93.78%. The experimental results show that the average accuracy of this algorithm is improved compared with single channel convolutional neural network and other literature algorithms, which shows that this algorithm can effectively improve the performance of EEG emotion recognition.

Abstract:To solve the problem that the posture of the double hemisphere capsule robot (DHCR) was likely to deviate from the targeted orientation due to nonlinear factors such as viscoelastic damping of the gastrointestinal (GI) tract and DHCR centroid deviation, as well as posture estimation error caused by large visual disparity and motion blur, a self-supervised learning-based rapid posture correction method was proposed. In terms of calibration of the capsule′s initial posture, the influence of the initial rotation angle was eliminated; in terms of estimation of the capsule′s attitude, based on spatial attention block(SAB) and temporal attention module (TAM), a spatial-temporal attention mechanism (TSAM) was designed by replacing a component of the standard convolution with the depthwise separable convolution and embedding it into PoseNet to generate an attention posture estimation network (APEN), which enhanced the model′s ability to extract features. The experimental results show that APEN can increase relative posture estimate accuracy by 52% while keeping inference speed almost unchanged when compared to the current capsule posture estimation method, Endo-SfM. Moreover, the accuracy of posture control is increased by 38.8% with this method, and it can correct the capsule posture in real time, laying the foundation for successful dynamic GI tract diagnosis and therapy.

Abstract:In view of the accuracy of the position stabilization time of industrial robots, enterprises and inspection institutions have been constantly studying and improving the detection methods. Through the comparative experiment of the laser tracker and the cable CompuGauge, the uncertainty measurement model is built, and the measurement repeatability, digital resolution, indication error and coordinate alignment are analyzed and evaluated using Class A and B methods. The expanded uncertainty of the laser tracker and CompuGauge are U1≈0.11 s and U2≈0.19 s respectively, and finally laser tracker；CompuGauge；position stabilization time；expanded uncertainty；|En|≈0.5. The experimental results show that the results of the laser tracker are more concentrated, stable, and more accurate, but are most affected by coordinate alignment, while CompuGauge is easy to operate, time-saving, independent of the robot model, but most affected by indication error.

Abstract:In view of the high complexity of the current high-precision load identification algorithm, which is difficult to achieve low-cost appliances monitoring devices, and the problem that it is difficult to achieve high-precision appliances identification only based on steady-state characteristics, designs of a low-cost appliances analyzing and monitoring device based on CUSUM and weighted Euclidean distance is proposed. The device collects the electrical parameter characteristics of the working state of the electrical appliance through the sampling circuit, establishes a three-dimensional electrical feature library based on the steady-state active power, reactive power and transient overshoot power amplitude, uses the CUSUM control chart to realize the detection of appliance switching event with the overshoot amplitude detection function, and completes the identification of appliance working state through the weighted European distance model. The experimental results show that the average appliance identification accuracy of the monitoring device designed in this paper is 97.3%, which verifies the feasibility of the design scheme.

Abstract:To address the problems of low sampling efficiency, poor environmental adaptation, and poor decision making that reinforcement learning faces in solving endtoend autonomous driving behavioral decision problems, a recurrent proximal policy optimization (RPPO) algorithm is proposed, which introduces a mobile inverted bottleneck convolution module and LSTM to construct a policy network and a value network, which effectively integrate the correlation information of front and back frames to achieve the prediction of multivariate situations by the intelligent body, improve the rapid cognitive ability of the intelligent body to the environment, and add L2 regularization layer to the value network to further improve the generalization ability of the algorithm, and finally manually set the intelligent body to keep the action constant in two consecutive frames, introduce a priori knowledge to constrain the search space and accelerate the convergence of the algorithm. Through CARLA open source simulation environment testing, the improved method significantly dominated the reward curve compared with the traditional method, and the success rates of three types of tasks, namely, straight ahead, turning, and designated route driving, increased by 10%, 16%, and 30%, respectively, proving that the proposed method is more effective.

Abstract:A UAV aerial object detection algorithm based on modified YOLOv4 is proposed in order to address the high need of detection speed for UAV aerial object detection as well as the issue of missed detection and false detection when there are many small targets in aerial images. Firstly, the lightweight network MobileNetv3 is introduced to replace the main feature extraction network of YOLOv4, and the depth separable convolution is used to replace the 3×3 standard convolution of the network, which reduces the complexity of the model and improves the detection speed. Secondly, the 104×104 shallow detection layer for small targets is added, the three detection scales of the original feature extraction network are increased to four, and the number of feature fusion network layers is increased. These changes increase the algorithm′s accuracy in detecting small targets. Finally, the K-means++clustering technique is used to redesign the initial anchor frame, accelerating the network′s rate of convergence. The UAV aerial data set is used in a comparison experiment. The findings demonstrate that as compared to the original approach, the suggested technique not only ensures average detection accuracy but significantly enhances the detection accuracy of small targets. The detection time is 15.2% faster while the model parameters are lowered by 60%. It performs accurately and with good real-time performance.

Abstract:In order to improve the state estimation of UAV in a large range of weak texture scenes, an improved visual inertial odometer combined with GPS positioning method is proposed. Firstly, the geometric structure information of the environment was represented by adding line features into the visual inertial odometer to improve the accuracy of pose estimation. Secondly, by introducing length threshold screening, the short line segments that do not contribute much to pose estimation are eliminated to improve the robustness of feature tracking. Finally, the GPS measurement information is fused with the improved visual inertial odometer in a nonlinear optimization way to correct the cumulative error of the visual inertial odometer. The simulation experiment based on EuRoC dataset and the real scene experiment applied to UAV show that, compared with the original algorithm, the positioning error of the line feature algorithm is reduced by 39.14% in the simulation experiment, 23.48% in the indoor scene and 33.58% in the outdoor scene. The point and line feature algorithm integrated with GPS. The positioning error was reduced by 53.99%.

Abstract:In order to solve the problems of slow response and instability in the flight and grasping process of UAV with manipulator in complex environment, the ISMO method is applied to the control of UAV with manipulator for the first time. Firstly, according to the position and attitude relationship in space and Euler-Lagrange equation, the overall kinematics and dynamics model of UAV and 3-DOF manipulator is established to ensure the accuracy of the system. Secondly, after using the mathematical model to describe the relationship between variables, a complex simulation environment is built to simulate the whole sampling process. Finally, the ISMO control rate is designed for the whole control, which is proved by Lyapunov equation. Considering the influence of the global dynamic and static environment on the disturbance of the UAV position, attitude, manipulator and the increase of load after grasping. The simulation results show that its response speed and robustness are better than the traditional PID controller, which ensures the efficient and stable operation of the system.