Abstract:In view of the urgent demand of small satellite for the flexibility of data acquisition, storage and down transmission, a control system of satellite solid state memory based on FPGA is proposed. The control system can realize the real-time storage and down transmission of small satellite in orbit data. Based on Verilog HDL language, through the top-down RTL level design, the full orbit telemetry and GPS data can be stored in real time, and the specific address segment data in NAND Flash can be transmitted down through the upper note instruction, so as to carry out multi angle analysis and mining on the ground. After the on orbit verification of multiple satellite models, batch production has been achieved. The fixed storage control system can continuously store 48.8 day delay telemetry and 45.2 day GPS data. On the basis of satisfying the data throughput rate, the bit error rate is as low as 8.16×10-9. By adding arbitration, state timeout, bad block management and three mode functions, the system has super high robustness, stability and single particle characteristics. The system fully meets the application requirements of the new generation of micro satellites for fixed storage and smart carrying.

Abstract:To a RSS visible light positioning system based on an amplifier with fixed gain, the detectable range of the light intensity is small, which leads to very low positioning accuracy in the area with weak signal and then results in a decreased space of effective positioning. A RSS visible light positioning system based on an amplifier with dynamically adjustable gain is proposed, where the gain of the main amplifier is adjusted automatically in light of the feedback from the output of the signal processor, thus the detectable range of the light intensity is extended, the positioning accuracy in the weak signal area is improved and the effective positioning space is expanded. Experiments show that within experimental space of 1 m×1 m×1.89 m, the positioning with high precision can be reached in 1.0 m×1.0 m×0.6 m space range, where the effective positioning space is expanded approximately 86.72% in comparison to a fixed gain positioning system. Experiments in a two dimensional plane indicate that in the coverage with low signal intensity, the average positioning error is decreased 47.05% for the dynamic gain system compared with the fixed gain system. The average system error is lowered to be less than 5 cm. The study leads to the conclusion that the positioning space is expanded effectively and the positioning accuracy in the weak light signal region is enhanced while the predefined high accuracy requirement of positioning is promised.

Abstract:Aiming at the problems of low accuracy and easy drift of the laser odometry constructed by traditional 3D point cloud registration algorithm in complex environments, this paper proposes an adaptive laser odometry for complex environments. First, the original point cloud data was collected by 3D Lidar, and after the point cloud preprocessing, the ground segmentation method was used to complete the point cloud data segmentation and obtain the road point cloud richness information; then, the NDT algorithm was used to convert the front and rear the frame point cloud data is zoomed to the maximum extent to realize the rough registration of the point cloud data; finally, under the guidance of the environmental judgment conclusion, the appropriate ICP algorithm was selected to complete the high-precision registration of the 3D point cloud and according to the output point cloud transformation relationship built the laser odometry. Through the data set and a large number of real vehicle tests in different environments, it is concluded that the average displacement error of the laser odometry in the indoor structured environment is 0.026 m, and the average displacement error in the outdoor unstructured environment is 0.1 m. The results show that the laser odometry constructed in this paper can better adapt to complex environments and obtain more accurate 3D point cloud maps and SLAM trajectories.

Abstract:A damage identification system of pipeline structure based on state space model is proposed for avoiding accident and real-time monitoring in this paper. This scientific and effective system is based on vibration signals to detect failures and faults in piping systems and aimed to comprehensively grasp the health status of piping system and give early warning. The results show that this system can response immediately in some special case, and the average response time of 4.436 s and the damage identification accuracy rate of 80% meet the design requirements, the response time is less than 20 s and the accuracy rate is more than 70%, respectively. This system can be adopted to various kinds of pipelines and deal with one or more failure mode thanks to the adaptation of fundamental theory and structure.

Abstract:As the core component of fiber optic gyroscope (FOG), the winding quality of the fiber optic coils (FOC) is critical to the accuracy of the FOG. In order to ensure the accuracy and efficiency of the fiber winding system, a defect detection method based on the improved YOLO algorithm is proposed. The model uses the Shufflenetv2 network to replace the convolution layer and pooling layer in the YOLO backbone network, which improves the feature extraction ability of the network; the Focus module is added to improve the training speed; the K-means clustering algorithm is used to cluster the original anchor boxes, and obtain a prediction frame suitable for fiber winding defect detection, the accuracy of defect detection is improved; at the same time, the loss function is modified, the CIOU is used to calculate the positioning loss, and the Focal Loss is used as the confidence loss and classification loss function to speed up the network convergence; and data enhancement is carried out to enhance the generalization ability of the network. It is concluded from the experiments that our proposed method is able to detect FOC winding defects with an average accuracy of 99.63%, which is an improvement of 2.06% over the original YOLO algorithm, and a detection speed of 91 fps. This will guarantee the practical application of the FOC winding system.

Abstract:The kernel entropy component analysis(KECA) feature extraction process only retains the maximum Renyi entropy information of the data, but does not fully utilize the category information of the data. As a supervised learning algorithm, linear discriminant analysis(LDA)can effectively extract category information in features. Therefore, a feature extraction method of KECA-LDA (KEDA) was proposed. Firstly, KECA reduced the dimension of the data according to the minimum Renyi entropy loss strategy. Then, the LDA algorithm was used in the KECA feature space to obtain low-dimensional features with discriminative information and input them into the support vector machine(SVM)classifier. The best performance SVM was obtained by beetle antennae search(BAS)and to build a fault diagnosis model. The KEDA-BAS-SVM method was applied to the Tennessee eastman(TE) for simulation experiments. The results showed that When the matrix similarity optimization based on distance measure was used to determine that the kernel parameter of RBF selected in KECA, compared with the KECA and LDA algorithms, after using KEDA feature extraction, the accuracy of multi-type fault diagnosis reached 99.7%, which verified the superiority of KEDA-BAS-SVM in the field of multi-type fault diagnosis.

Abstract:The short-term load forecasting model using long short-term memory(LSTM) network has the problem of feature redundancy and loss of important information. In order to solve these problems, a shortterm load forecasting method based on feature selection strategy and DLSTMsFCN is proposed. Firstly, the feature optimization strategy based on extreme gradient boosting(Xgboost) is adopted to improve the feature redundancy problem of load prediction input. Secondly, DLSTMs are used to extract the time series features of load data, and the highresolution information is extracted through the multidimensional convolution operation of FCN and structural features. The purpose is to enhance the learning and memory of important features of input data, and then form an efficient and accurate shortterm load forecasting model in parallel. The experimental results show that compared with ALSTMs and CNNLSTMs, the prediction error of the optimization method in this paper decreases by 6% and 4% respectively, and the prediction error fluctuation decreases by 4.7% and 4.8% respectively.

Abstract:Aiming at the problems of how to run the microgrid cluster stably and how to reduce the operation cost of the microgrid cluster to the greatest extent, the energy optimization of the microgrid cluster based on the improved bat algorithm was studied. The system takes the microgrid cluster composed of two AC microgrids and one DC microgrid operating off grid as the research object, constructs the microgrid cluster architecture including centralized energy storage system, establishes the optimal scheduling model using the operation cost and environmental impact cost as the multiobjective function, and transforms the multiobjective function into a single objective function through the binary comparison weighting method, The improved bat algorithm based on tent mapping and Cauchy variation is used for energy optimization. The results show that compared with the traditional microgrid cluster optimization method, the system energy optimization method has better system operation stability, effectively improves the economic efficiency of the system, and reduces the comprehensive daily operation cost by 14.63%.

Abstract:In ultra-wideband indoor positioning, the signal propagates in the non-line-of-sight scene due to the influence of various obstacles in the complex indoor environment, resulting in positioning errors. Aiming at the issue of the influence of nonlineofsight propagation on indoor positioning accuracy, a nonlineofsight recognition method based on improved complete ensemble empirical mode decomposition with ICEEMDAN was proposed. Firstly, the channel impulse response is completely decomposed to obtain the IMF with different scale characteristics. Secondly, the Pearson correlation coefficient method is used to select some IMFs for reconstruction to retain more effective information, and the wavelet transform is used to obtain effective timefrequency characteristics of the reconstructed signal. Finally, the nonlineofsight signal is identified by constructing a convolutional neural network. The experimental data are based on the 802154a UWB model and opensource data set. The experimental results indicate that the average accuracy of the proposed recognition method reaches 98.5%, which is 5.6% higher than that of other algorithms in the simulation data set and 14.3% higher than that in the PDS data set, which verifies the effectiveness of the proposed recognition method.

Abstract:Aiming at the problem that point clouds are prone to misregistration under multiple interference conditions of noise, occlusion and similar features, a point cloud registration method based on multifeature key point extraction algorithm and adaptive scale fusion features was proposed. In keypoint extraction, multiple features are computed simultaneously to make keypoints more descriptive and robust. On the basis of adaptive scale, FPFH and RoPs features are used for initial registration and error point pair elimination, and multiple similar transformation matrices are obtained respectively. After the above solution is completed, the matrix obtained by the two is formed into a set for clustering, and the class with the largest number of matrices is averaged as the final result to complete the feature fusion. The experimental results show that the RMSE, ATI and ERR of the proposed algorithm are 046 mm, 1 and 037 in the actual scenario when a few unrealized error registration points are ignored. The accuracy of the dataset test was 99.3%. It shows that the algorithm has high accuracy and robustness.

Abstract:An indoor dynamic positioning algorithm based on UWB and inertial measurement unit (IMU) fusion is proposed to deal with the problem that ultra-wideband (UWB) positioning is susceptible to various noises and nonlineofsight (NLOS). The algorithm firstly uses the extended Kalman filtering algorithm to filter the position information based on the angle of arrival (AOA) positioning method, and synchronizes the time with IMU data. By comparing the change speed of UWB position information at adjacent times with the movement speed of tags measured by IMU, the algorithm realizes the recognition and compensation of NLOS data, thus reducing the impact of NLOS on positioning accuracy. Then the improved particle filtering algorithm is used to optimally estimate the fused data to suppress noise interference and finally achieve accurate label location. The experimental results show that the proposed algorithm using AOA based location method can save the hardware cost while ensuring the location accuracy. Compared with the positioning scheme using only UWB sensors, the proposed algorithm can effectively reduce the positioning error of UWB according to the prior information provided by IMU, and has high reliability in the nonline of sight environment. Contrary to the fusion algorithm based on extended Kalman filter and unscented Kalman filter, the positioning accuracy is improved by 656% and 560% respectively. In contrast to the standard particle filtering algorithm, the running time of the proposed algorithm based on the improved particle filtering algorithm is reduced by 423%.

Abstract:n order to reduce the labor and timeintensive burden of manually marking arteries and veins, an algorithm for automatic segmentation of retinal arteries and veins based on the ResNet_UNet network model is proposed in this research. First, retinal images were acquired using a multispectral retinal imaging system and a dataset was made. The dataset contained 206 retinal fundus images at 548 nm wavelength and their pixellevel labels. Then, the multiscale feature extraction module and loss function module in the ResNet_UNet network model were optimized. And a channel attention mechanism and postprocessing methods were added to improve the accuracy of automatic classification of arteries and veins. Finally, 165 images were randomly selected from the dataset as the training set, and 41 images were tested as the test set. Experiments show that the deep learning model established in this study can automatically and accurately segment the arteries and veins in retinal images, with an accuracy rate of 98.50%. Keywords：

Abstract:In order to suppress the background interference in the complex environment of the openpit mine and perform accurate semantic segmentation of the safety retaining wall, a segmentation method of the safety retaining wall of the openpit mine based on the improved DeepLabV3+ network is proposed. First, the backbone network adopts the lightweight MobileNetV2 network, which effectively reduces the amount of network parameters and computation through the depthwise separable convolution and inverted residual structure. Then, a hybrid attention module is added for channel and spatial feature enhancement, which can avoid the loss of edge information. Finally, data augmentation and transfer learning are used to solve problems with fewer target datasets and improve the generalization ability of the model. The experimental results show that the method has a good segmentation effect, with MIOU and MPA of 8506% and 9294%, respectively, which are better than the original network and other classic network models.

Abstract:Prohibited items detection in X-ray security inspection is widely used to maintain public traffic safety and personal safety. In order to solve the problems of variable shape and scale, severe overlap and occlusion in Xray images, an improved YOLOv5s model combining deformable convolution and attention mechanism is proposed for prohibited items detection. Firstly, deformable convolution is introduced into the backbone network to enhance spatial feature information extraction by learning sampling offsets to adapt to different deformations of objects. Secondly, the mixed convolution attention module is used to enhance the model’s ability to perceive the detected target and suppress irrelevant background interference. Then a channelguided atrous space pyramid module is constructed to obtain more accurate global contextual information and improve the model′s ability to identify overlapping occlusion targets. Finally, the CARAFE operator is used to replace the nearest neighbor interpolation to make full use of the content information in the upsampling process and improve model’s detection accuracy. The experimental results on the SIXray_OD and OPIXray datasets show that the model’s mAP@05 is 21% and 18% higher than the original YOLOv5s, reaching 906% and 900%, respectively. Compared with many existing advanced algorithms, it has better detection accuracy and realtime performance.

Abstract:Aiming at the problem of limited computing power and storage space of edge computing devices, a lightweight human fall detection algorithm based on YOLOX was proposed. Firstly, Ghost module in GhostNet is used to reduce the redundancy of convolution parameters in Neck and Prediction layers in YOLOX. Secondly, the coordinate attention mechanism is added to the Neck layer to enhance the ability of key information extraction and reduce the influence of background noise. Finally, aiming at the problem of insufficient detection ability of lightweight model detection head, an auxiliary head module is introduced to strengthen the learning ability of lightweight detection head. The performance of the proposed model is tested by the algorithm and the experimental results are run on the edge computing end of NVIDIA Jetson Xavier NX. The results show that the mAP@05 of the proposed model reaches 849%, and the model size is 256 MB. Compared with the YOLOX model, only a small amount of reasoning speed is sacrificed to improve the detection accuracy of 4.6% and reduce the model size of 25.4%. In addition, compared with some mainstream object detection algorithms, the proposed model also has certain advantages. These results show that the proposed model can better meet the requirements of lightweight and accuracy of the model for edge computing equipment in human fall detection.

Abstract:Aerial insulator images are characterized by complex background, variable spatial direction of insulators and large aspect ratio. A large amount of background information will be introduced to detect aerial insulators using the traditional horizontal rectangular box object detection algorithm, which makes it more difficult to locate insulator defects and other subsequent operations. Therefore, this paper proposes an arbitrary angle aerial insulator detection method based on neighborhood interactive attention. The method in this paper adds the angle prediction branch of the object bounding box to the existing insulator detection framework. The introduction of angle prediction branch makes the method put forward higher requirements for network feature expression ability. Therefore, a neighborhood information interactive attention mechanism is proposed in this paper, and the spatial information relationship of insulators is modeled through this mechanism, so as to achieve accurate detection of aerial insulators. The test results show that the proposed method can accurately detect insulators in complex environments, and the detection accuracy reaches 935%, which effectively avoids the interference of a large amount of background information on insulator defect detection.

Abstract:Aiming at the problems of low detection accuracy and slow speed of the traditional method for detecting surface cracks on ballastless track slabs of high-speed railways, an improved CenterNetbased algorithm for detecting surface cracks on track slabs is proposed. The algorithm adds atrous space pyramid pooling module (ASPP) between the codec network as a way to expand the perceptual field of the feature map and fully extract the contextual information at different scales. Then adds a multispectral channel attention module (MCA) to the feature extraction network so that the network can better learn the weights of each channel and capture the image rich input feature information. Finally, the αIoU loss function is used to improve the accuracy of bounding box prediction. The experimental results show that the mean average precision(mAP) of the proposed algorithm reaches 8412%, which is 337% higher than that of the traditional algorithm, and it has a good detection effect on the surface cracks of the track plate.

Abstract:In order to achieve intelligent rapid detection of pavement defects, the deep learning object detection algorithm YOLOv5 is improved, and the three detection models （YOLOv5A, YOLOv5C, YOLOv5AC）can be quickly detected by video detection. Using smart phones and digital cameras to collect road defect images and make data sets, to meet the needs of video detection, the use of Kmeans algorithm and 1IoU as sample distance recluster anchor, to obtain better anchor frame parameters; the introduction of CBAM attention mechanism in multiple structures of the network, enhance the feature extraction ability of the model. The experimental results show that the average accuracy of the YOLOv5C algorithm on the training set reaches 918%, which is 1% higher than that of the original model. The average accuracy of the YOLOv5A algorithm on the verification set reaches 927%, which is 17% higher than that of the original model. In terms of actual detection effect, the YOLOv5AC algorithm achieves 89%, 62%, and 90% in the identification accuracy of cracks, broken plates and pits, which is 45%, 4%, and 5% higher than the original model. And the detection speed of the model reaches 40 FPS. YOLOv5AC algorithm has high detection accuracy and recognition speed, and can meet the intelligent realtime detection requirements in road defect detection under certain conditions.

Abstract:Remote sensing image superresolution reconstruction based on Generative adversarial networks has some problems, such as unstable training, redundant parameters and unclear texture details. This paper presents a super resolution reconstruction algorithm of remote sensing image based on edge detection. Firstly, the improved Canny edge detection operator is introduced into the generator network for lowresolution image feature extraction. Bilateral filtering and 3×3 neighborhood gradient are used to detect image edge information in the Canny operator edge extraction process, so that the network can better express highfrequency features. Secondly, in order to reduce the network parameters and improve the stability of network training, the redundant BN layer in the discriminator network is removed, and the Wasserstein distance is defined as adversarial loss to solve the gradient disappearance phenomenon in generating adversarial network training. On the NWPU RESISC45 dataset, Compared with WDSR and CARN, the peak signaltonoise ratio and structural similarity of the proposed method are improved by 1.22 dB,0.114 and 0.32 dB,0.013, respectively. Moreover, compared with other SR algorithms such as WDSR and CARN, the reconstructed images are improved in texture details and subjective visual effects.

Abstract:Lane detection is a significant research subject in the field of intelligent driving. However, there will always be inaccurate lane segmentation and insufficient realtime processing capabilities in practical applications. Accordingly, an improved algorithm based on the Pyramid Scene Parsing Network is proposed. A main network PSPNet is built on a basis of the encoding structure, and the encoder backbone network is replaced by the lightweight MobileNet v2 network, which effectively cut down the parameter amount and computational complexity of the whole network. Hole convolution is added into the network and feature fusion is realized between different layers, which expands the model receptive field and enriches local feature information. Finally, an adaptive line fitting algorithm is used to fit different lane lines in order to obtain the final prediction result. The Caltech lane data set is come into use for testing. The experimental results show that the improved algorithm has better segmentation for different types of lane lines. Compared with the original algorithm, the Pixel Accuracy and the Intersection over Union is improved by 3.91%, 4.14%, and FPS up to 28 frames per second. The segmentation accuracy and inference speed of the proposed algorithm are superior to other comparison algorithms.

Abstract:Buildings in aerial images often have multiple orientations. The target detection algorithm based on the traditional convolutional neural network mostly uses the horizontal anchor frame as the detection frame, which has a low accuracy in detecting the building scene with multi orientation distribution. Therefore, this paper proposes a target detection algorithm based on CenterNet neural network, adds angle prediction branch on the basis of CenterNet model, and integrates the orientation angle information into the network. Aiming at the problem that few building angle features are extracted in the feature extraction stage of CenterNet model, asymmetric convolution is used to replace the original 3×3 convolution to enhance the feature extraction ability of neural network for rotating target angle information, and reduce the impact of angle periodicity on target detection by improving the loss function. The improved network can more accurately detect buildings with multi orientation distribution. In this paper, experimental tests are carried out on the data set built by ourselves. Under the same environment, the overall average precision is improved by 5.2% before and after the network improvement，including 74% for buildings with large orientation changes within the range of 10°~80° and 100°~170°. The average precision of buildings with small orientation changes within the range of 0°~10°, 80°~100° and 170°~180° has increased by 31%, effectively improving the average precision of buildings with multiple orientations.

Abstract:When the OFDM system uses the RootMUSIC algorithm to complete the joint TOA and DOA estimation, the required roots appeared in the form of conjugating symmetry which will be computational redundancy. Aiming at this problem, a RootMUSIC algorithm based on spectral factorization—SFRootMUSIC algorithm is proposed. Based on the structural characteristics of Laurent polynomials, the algorithm uses spectral decomposition to reduce the order for the root polynomial by a half, which reduces the computational complexity, completes independent delay and angle estimation, and constructs a cost function of parameter pairing, complete the joint estimation. The simulation results show that the SFRootMUSIC algorithm has similar estimation performance with the RootMUSIC algorithm, and its complexity is lower. When the number of array elements is 12,the number of subcarriers is 512 and the number of snapshots is 512, the complexity can be reduced by 6994%. The proposed algorithm can achieve the joint estimation of TOA and DOA with lower complexity while ensuring the accuracy, which verifies the proposed algorithm is more suitable for realtime computing.

Abstract:To address the problem that the bandwidth and gain of conventional analog equalizers cannot be satisfied simultaneously, inductive cresting technique and negative impedance converter are used for high frequency operation. The bandwidth is extended by using active inductors to achieve inductive cresting techniques, while a negative impedance converter consisting of crosscoupled transistors converts capacitance into negative capacitance to offset the output capacitance, increasing peak gain while maintaining DC gain. The circuit is designed based on the SMIC 130 nm process library. The simulation results show that the equalizer works under the power supply voltage of 18 V, which can well compensate the severely lost channel in the highspeed transmission system at the rate of 6375 Gbps. The horizontal opening of the eye diagram reaches 085 UI.

Abstract:Aiming at the difficulty of node replacement and limited energy of underwater wireless sensor network, the paper proposes an energy-efficient routing protocol based on improved sine cosine algorithm. In the process of cluster formation, it uses the improved sine cosine optimization algorithm to select the cluster head, considering the three factors of energy, node density and communication distance, and discusses the corresponding weights, then designs a more reasonable fitness function for cluster head selection. Singlehop transmission within the cluster, between the clusters used multihop transmission. The algorithm transmits the collected information to water surface by limiting the depth, energy and forwarding area to select the appropriate next hop. The network simulation results show that under same conditions, the work hours increases by 6910%,2478% and 1494% compared with the traditional LEACH algorithm, the KACO and DUCISCA algorithm, which can effectively balance network energy consumption, prolong network life, and improve data transmission rate.

Abstract:The channel measurement based on pseudo-random sequence was carried out for 5.6 GHz Multiple Input Multiple Output (MIMO) system in Shanghai metro tunnel environment. The impulse response of subchannel is obtained by correlation operation. The effects of antenna polarization and array element spacing on the condition number, equivalent degree of freedom and channel capacity of MIMO channels are analyzed. The results show that when the average number of cross polarization conditions of the antenna is about 10, it is about 5 and 8 lower than vertical polarization and horizontal co polarization respectively. When the mean value of cross polarization equivalent degree of freedom is about 16, it is about 03 and 02 higher than that of vertical polarization and horizontal homopolarization respectively. When the average cross polarization capacity is about 58 bit/s/Hz, it is about 07 bit/s/Hz and 08 bit/s/Hz higher than that of vertical polarization and horizontal polarization respectively. All three results agree that the cross polarization of antenna is beneficial to the improvement of MIMO performance. It is also found that the antenna array element spacing has a greater impact on the channel capacity of co polarized MIMO. The channel capacity of cross polarized MIMO is less affected. It shows that cross polarization can obtain larger channel capacity with smaller array element spacing. These results have reference significance for reducing antenna size and increasing channel capacity in narrow space of subway tunnel.

Abstract:Minimum bending radius needs to be strictly controlled when laying cables. Accurate segmentation of cable laying image is the basis of controlling bending radius. Traditional visual and classical semantics segmentation methods do not work well for target segmentation of long and thin cables in complex environment. This paper presents a new cable semantics segmentation method based on improved dualmode fusion semantics for ESANet network. Instead of the RGBD Fusion module in ESANet, an efficient SAGate is used to complete the dualmode feature correction and fusion tasks. The fused features participate in the feature extraction of the subsequent two modes at the same time to achieve accurate segmentation of the thin feature cable mask. By collecting RGB and corresponding depth images of cables with different postures, the results show that the improved ESANet network has a good segmentation effect on slender feature targets such as cables, which is 399% higher than Net model segmentation accuracy (mIoU), and 7.68% higher than SwiftNet singlemode semantics segmentation network of RGB. This method can be extended to other target segmentation tasks with slender feature.

Abstract:Due to insufficient consideration of various motion orientations in Merge mode, reducing the prediction accuracy of versatile video coding (VVC) interframe prediction, a VVCbased multidirection refinement Merge mode optimization algorithm is proposed. Based on the analysis of Merge mode with motion vector difference, the algorithm first adjusts the step size selection range and adds multiple search directions, then adaptively selects the motion compensation method of chroma block according to the step size. And finally selects the optimal motion vector information according to the ratedistortion cost criterion. The results show that, comparing with VTM120 reference model, the BDrate of Y, U and V components decreases by 057%, 062% and 025% in lowdelay Pframe configuration, respectively. In the random access configuration, the Y component of BDrate decreases by 027%, the U component only increases by 011%, and the V component decreases by 004%, which effectively improves the coding performance of Merge mode.