2021, 30(11):3-10. DOI: 10.15888/j.cnki.csa.008346
Abstract:With decades of development, X86 and ARM have gradually dominated the markets of desktops and mobile phones. Although these two architectures are becoming increasingly powerful from the standing points of technical advances and software ecosystem, they are not good candidates for architectural research due to their complicated Instruction Set Architecture (ISA) definitions, comprehensive technical designs and intimidating copyright protection issues. Before the introduction of the open RISC-V ISA, there is no appropriate ISA for computer architectural research and innovation. RISC-V has attracted attention and participation from both the industry and academia. Hardware Performance Counter (HPC) is an important tool for researching and optimizing computer processor cores. The original definitions of HPC in the RISC-V standard do not scale properly and the number of events simultaneously monitorable is rather small. For these reasons, we propose a new distributed HPC based on RISC-V in this study. We have integrated this design into the lowRISC-v0.4 open SoC platform and run it on the Genesys2 FPGA board. Our HPC only uses three Control and Status Registers (CSRs) to capture all events. The number of events that can be concurrently monitored is one order of magnitude higher than that the RISC-V standard can support. Meanwhile, our strategy could provide detailed and accurate data for researchers focusing on the performance analysis of RISC-V processors, the architectural optimization, and side-channel attack and defense.
2021, 30(11):11-19. DOI: 10.15888/j.cnki.csa.008351
Abstract:In computer systems, the memory overflow attack is a long-existing security problem and is still common nowadays, which can be effectively hindered by pointer encryption. Nevertheless, the implementation of the technique by software significantly lowers the program running efficiency and leads to additional memory overhead. In this study, we develop an encrypted/decrypted pointer coprocessor PEC-V based on the Rocket Custom Coprocessor (RoCC) interface of RocketChip. The overflow attack can be prevented through the control of encryption/decryption of the return address and function pointer by the coprocessor under the user-defined instruction of RISC-V. PEC-V mainly depends on Physical Unclonable Function (PUF) to avoid storing the key value of the encrypted pointer in memory. Thus, this mechanism not only ensures the randomness of the key value, but also reduces the times of accessing memory. The experimental results show that PEC-V is defensive against various buffer overflow attacks while the program running efficiency is only reduced by approximately 3% on average, which is better than previous mechanisms.
2021, 30(11):20-26. DOI: 10.15888/j.cnki.csa.008347
Abstract:The RISC-V instruction set architecture features modularization and scalability. On the basis of the integer instruction set, the RISC-V architecture based processors can optionally support the official standard and non-standard custom instruction set extensions. This also means that, for each new custom extended instruction set, users need to implement corresponding support in the compiler toolchain. After analyzing the LLVM compilation framework and researching the general methods supporting RISC-V custom extended instructions, we conduct the implementation and verification with the XuanTie C910 custom instruction set as an example. The results can provide references for the research and implementation of RISC-V custom instruction set extension based on LLVM infrastructure.
2021, 30(11):27-32. DOI: 10.15888/j.cnki.csa.008349
Abstract:For correct and optimized machine instructions, it is necessary to design and use a suitable program stack frame layout during the code generation stage of the compiler back-end. Due to the scalability of the RISC-V vector extension architecture and the unknown length of its vector register at compile time, the traditional stack frame layout cannot be applied. Although the previous stack frame layout implemented for vector extension in LLVM can generate correct machine instructions, it has problems such as many load/store instructions and reserved registers as well as large stack frame sizes. We analyze the problems existing in the previous implementation and propose a new layout and vector object calculation method on this basis. Then we verify it through the test set developed by the Barcelona Supercomputing Center. Experiments show that the new stack frame layout can greatly reduce the number of load/store instructions and stack space.
2021, 30(11):33-40. DOI: 10.15888/j.cnki.csa.008348
Abstract:In this study, we design and implement an automatic testing system for semantic equivalence of RISC-V assembly programs. While developing RISC-V programs, especially developing efficient programs based on extension instructions (such as vector extension), developers often write assembly code manually. For example, for the standard C function library, we often write the corresponding vector version functions for better performance. Without the compiler, the manually developed assembly code can maximize the efficiency of the program, but it skips many important compilation processes (such as type checking and register allocation), thus putting forward higher requirements for the developers. It will greatly affect the correctness of the code and the efficiency of software development and debugging if we can quickly and automatically test whether the rewritten version is semantically equivalent to the standard version of the program. The existing RISC-V testing framework lacks support for semantic equivalence testing and fails to consider the side effects caused by program executions. Based on the dynamic test environment of a simulator, this research designs and implements an automatic testing system for semantic equivalence of RISC-V assembly programs. It can capture side effects caused by program executions through monitoring machine states and generate testing reports with user-defined testing targets. Experiments show that the system, compared with existing testing systems, can test the semantic equivalence of RISC-V assembly programs.
2021, 30(11):41-45. DOI: 10.15888/j.cnki.csa.008350
Abstract:This study introduces the hardware connection scheme for an embedded intelligent car control system based on RISC-V, the state analysis method of an intelligent car based on a state machine, and the motor control scheme in different application scenarios. The system takes the FPGA development board running the RISC-V softcore as the main control board of the intelligent car and collects the signals from the ultrasonic sensor and infrared sensor of the intelligent car through the GPIO module of RISC-V to detect the obstacles in front and rear of the car respectively. Moreover, it uses the GPIO interrupt to respond quickly to the signals from the collision detection sensor and tilt angle sensor and adopts the PWM module for the motor control in different scenarios. The test results show that the control system introduced in this paper can fulfill the functions of the intelligent car, such as autonomous obstacle avoidance, collision detection, and attitude detection.
2021, 30(11):46-53. DOI: 10.15888/j.cnki.csa.008170
Abstract:Ranking entities according to the relevance degree between the given entity and other entities in a Knowledge Graph (KG) is critical for related entity search. The relevance between entities is not only reflected in the KG but also the rapidly generated Web documents. In existing methods, the relevance degree is mainly calculated from the KG, which cannot reflect the knowledge rapidly evolving in the real world, and thus effective results cannot be obtained. Therefore, in this study, we first propose an algorithm for searching candidate entities on the basis of the TransH model by analyzing the semantic representation of entities in hyperplanes of different relations. To improve the precision of ranking candidate entities, we propose an Entity Undirected Weighted Graph (EUWG) model by quantifying the relevance between searched and candidate entities reflected in Web documents and KG. Experimental results show that the proposed method can precisely search and rank the candidate entities in the large-scale KG.
2021, 30(11):54-62. DOI: 10.15888/j.cnki.csa.008180
Abstract:The task of document classification in natural language processing requires the model to extract high-level features from low-level word vectors. Generally, the feature extraction of deep neural networks uses all the words in the document, which is not well suited for documents with long content. In addition, training deep neural networks requires massive labeled data, which often fails to achieve satisfied results under weak supervision. To meet these challenges, this research proposes a method to deal with weakly-supervised long document classification. On the one hand, a small amount of seed information is used to generate pseudo-documents to enhance training data to deal with the situation where accuracy is difficult to improve due to the lack of labeled data. On the other hand, using recurrent local attention learning to extract summary features based on only a few document fragments is sufficient to support subsequent category prediction and improve the model’s speed and accuracy. Experiments show that the pseudo-document generation model can indeed enhance the training data, and the improvement in prediction accuracy is particularly significant under weak supervision. At the same time, the long document classification model based on the local attention mechanism performs significantly better than benchmark models in prediction accuracy and processing speed, with practical application value.
2021, 30(11):63-70. DOI: 10.15888/j.cnki.csa.008232
Abstract:With the outbreak of the COVID-19 pandemic, most colleges and universities have adopted VPN to ensure remote learning and scientific research when students and teachers cannot return to school. To understand the specific situation, we collect the VPN logs of Peking University during the COVID-19 pandemic from February 2020 to September 2020 and discuss the number of users, login and logout time, usage time, cluster analysis, and user categories. The maximum number of daily users of VPN is about 15 000, and the maximum number of concurrent users is about 5000. Moreover, the average daily usage time of VPN is up to 325 minutes. These data show that students and teachers rely highly on VPN for remote learning and scientific research. According to the daily average usage time and the number of days of use, the users can be roughly divided into 4 categories. The VPN usage time of science and engineering users is slightly longer than that of liberal arts users, but the trend of change is the same. These data are of reference value for understanding VPN usage and adjusting VPN resources. Although VPN has facilitated the acquisition of school resources during the COVID-19 pandemic, the security risks it brings cannot be ignored. If protections are not enough, the user terminal will be vulnerable to attacks by hackers, who can steal resources or attack other machines in school using the user terminal as a springboard. As remote learning and scientific research will become the new normal, is it “sweet” or “poisonous”? Colleges and universities should be well prepared to deal with it.
2021, 30(11):71-81. DOI: 10.15888/j.cnki.csa.008145
Abstract:People put forward a higher service demand for traditional ward care in the process of intelligent technology reforming various traditional industries. According to the field investigation of traditional wards, the self-care ability of patients should be improved in daily ward life and the real-time monitoring of their living conditions by medical staff as well as families is waiting to be strengthened. To this end, this study proposes a smart ward control system integrating speech and ElectroEncephaloGraphy (EEG) and combined with the Internet of Things (IoT), thereby realizing the control of infrastructures such as ward appliances and real-time monitoring of cloud wards. Furthermore, the EEG-based control method based on eye blinks can meet the higher requirement for a patient’s physical conditions. According to the two experiments involving 10 subjects, speech recognition achieved an accuracy of 98%, and the accuracy of EEG-based blink recognition for healthy people and patients is 94.3% and 82.9%. The results show that the system can operate stably in complex environments such as wards, providing patients with a more intelligent and comfortable convalescent environment.
2021, 30(11):82-90. DOI: 10.15888/j.cnki.csa.008183
Abstract:In China, the number of mobile Internet users has reached 932 million, accounting for 99.2% of the entire Internet population. Moreover, the penetration rate of smart phones among China’s college students is close to 100%, making it possible to use smart phone APPs to assist in teaching in colleges and universities. This work studies the method of introducing mobile APPs for learning evaluation and management in college classrooms, namely integrating traditional courses, class management, and some evaluation methods into mobile APPs, to improve the efficiency of teaching and management in classrooms and enhance students’ enthusiasm for learning knowledge. In this study, 884 students in 16 classes of three courses were involved in a mobile classroom test. Among them, 165 students in four classes of two courses participated in the contrast experiment between mobile classroom tests and paper tests, and 124 students were investigated with mobile classroom tests. The results show that compared with the traditional paper test, the mobile phone test gets lower scores but arouses great interest of students and bursts the class activity. In general, most students agree that the mobile phone test represents a tendency and may become an important evaluation means in future teaching. However, the mobile phone evaluation system in classrooms is still an emerging thing with strong vitality and requires further research.
2021, 30(11):91-98. DOI: 10.15888/j.cnki.csa.008148
Abstract:In the traditional industrial Internet platform, the security and privacy issues of data generated by terminal equipment are the main bottlenecks hindering the development of the industrial Internet. With the geometric growth of terminal data, protecting the security and integrity of data has become the core research area of the industrial Internet. This study first designs blockchain-based equipment and data management architecture, providing a reliable, tamper-proof database. Then, the digital certificate is employed to provide the authority access control mechanism for the platform to improve the security level of platform access. Secondly, the terminal equipment and its configuration files are managed indirectly through the chain code, which avoids the data pollution caused by the random access of the terminal equipment. Finally, the data generated by the terminal equipment is packaged and encrypted through the public and private keys of the equipment itself and then stored on the blockchain with the consensus mechanism of the blockchain. Experiments show that the proposed scheme has good stability, safety, and operability.
2021, 30(11):99-105. DOI: 10.15888/j.cnki.csa.008168
Abstract:For better performance of the projection mapping technology applied to the digital exhibition of museums and high-quality projection, a 3D digital display system for cultural relics is designed with realistic 3D modeling and a high-fidelity appearance. The system relies on the physical model of mechanical rotation and the image of real cultural relics and combines multiple projection mapping technologies with the optical see-through display. Both geometric and radiometric calibration methods are used to correctly project a high-quality texture onto the moving 3D projection surface without perspective distortion to increase the projection effect and the degree of freedom of the projection content. For the evaluation of the influence of the ambient light and the shadow occlusion by the surrounding objects in general museums on the system, the projection area and display area of the system are tested with different intensities of light. Experimental results show that the system has strong robustness to the influence of the ambient light and the potential for deployment in a general museum context, making the digital display of cultural relics richer.
2021, 30(11):106-111. DOI: 10.15888/j.cnki.csa.008130
Abstract:The paint repairing robot for cars needs to face vehicles in different sizes and adapt to a variety of modeling surfaces and colors. This highly adaptive requirement makes the design of the paint repairing robot far more difficult in target tracking, path planning, motion space, and other aspects than the painting robot in the automotive manufactory. Therefore, the painting path re-planning is necessary. First, a large amount of point cloud data is segmented by parts, and then the eight-neighborhood method is used to calculate the contour of the closed surface. Finally, the grating path is generated on the surface by the slicing method, and the critical path of each surface is formed. The eight-axis truss robot system is designed and manufactured, and the path planning of the eight-axis linkage is generated by the ant colony algorithm. Subsequently, the generated path and trapezoidal curve acceleration are sent to the PLC-based motion control program through the ADS protocol of the Beckhoff controller to complete the linkage and collaborative paint repairing movement of each joint axis. The experimental results show that the system can automatically control the axis of the robot tool to align the arbitrary surface with the normal vector for different vehicles and drive the eight axes to track the surface motion smoothly. The system can also be widely applied to the robot machining of various surfaces.
2021, 30(11):112-117. DOI: 10.15888/j.cnki.csa.008169
Abstract:The charity field is always faced with the problem that the data cannot be open and transparent. People cannot trust the charitable organizations that do not disclose the data. Even if the data is disclosed, they have to face the question of data fraud. Aiming at the problems existing in charitable organizations, such as lack of credibility, low convenience, and opaque flow of money, this system adopts the blockchain. By using a new data storage model, the uploaded data is encrypted and decrypted according to the required requirements, so that the data has confidentiality. At the same time, the data in the public chain can not be modified to make up for the transaction in the public chain At the same time, there may be insufficient data fraud.
2021, 30(11):118-126. DOI: 10.15888/j.cnki.csa.008177
Abstract:Tragedies caused by not wearing safety helmets occur from time to time in engineering sites. To assist site managers in protecting workers’ safety, this study has designed and implemented an intelligent safety helmet supervision system based on deep learning. The system adopts the YOLOv4 target detection model integrating speed and accuracy, generates the new anchor boxes by K-means clustering analysis on the data set, and then trains the model with the new anchor boxes. The detection accuracy of the helmet is improved to 92%. The detection model YOLOv4 is combined with the tracking model DeepSORT to effectively solve the problems of repeated warnings and failures to produce statistics on illegal data. Finally, it is built into a cross-platform mobile APP, which is convenient for managers to use the mobile terminal to monitor the helmet-wearing situation anytime and anywhere. This intelligent supervision system covers a set of functions including safety helmet detection, real-time broadcast of detection videos, intelligent warning, illegal picture capture and display, and illegal data visualization. It can greatly improve the production safety factors and supervision efficiency in the project site.
2021, 30(11):127-137. DOI: 10.15888/j.cnki.csa.008166
Abstract:This work studies a new practical combinatorial optimization problem (known as 2L-VRPB) which combines the classic Vehicle Routing Problem with Backhauls (VRPB) and two-dimensional Bin Packing Problem (2L-BPP). The 2L-VRPB aims to find the route set at the minimum cost for a homogeneous fleet of vehicles to satisfy the delivery requirements of linehaul customers and the pickup demands of backhaul customers. This study investigates two versions of the 2L-VRPB. Both versions are loaded with unrestricted loading, but one is packed with rotation while the other is not. These two variants are frequently employed in the industry of appliance maintenance service and grocery, but they have been less examined in the literature. To solve these two variants, we propose a metaheuristic integrating an enhanced memetic algorithm with a combinatorial packing heuristic. The packing algorithm checks the loading feasibility by employing five basic packing heuristics with two additional improvement strategies. Extensive computational experiments show that the proposed metaheuristic is a practical and effective solution to both VRPB and 2L-VRPB.
2021, 30(11):138-144. DOI: 10.15888/j.cnki.csa.008128
Abstract:Concerning the problem of poor detection of human body keypoints based on video streams and possible motion blur after video stream slicing, an improved RetinaNet-CPN network is proposed to detect the keypoints, avoiding the interference of motion-blurred images after slicing and improving the detection accuracy of the keypoints. After the video stream is sliced, the improved RetinaNet network is first used to detect all the people in the picture and perform fuzzy detection on each target frame. The target frame larger than the threshold is deblurred, and finally, the keypoints are extracted with the CPN network with the attention mechanism. After the IOU function of RetinaNet to measure the difference between the predicted frame and the real frame is changed into DIOU, the target detection AP increases by nearly 3% in the simulation experiment. For blurry pictures, the blur kernel estimated with the spectrum feature of uniform linear motion is slightly different from the actual blur kernel, and the original clear picture can be restored after the deblurring. At the same time, the attention mechanism is adopted to assign reasonable weights to each channel and feature layer, which increases the CPN detection AP by nearly 1% and the AR by 0.5%.
2021, 30(11):145-154. DOI: 10.15888/j.cnki.csa.008127
Abstract:We propose a fuzzing method based on function importance, because the existing fuzzing methods lack fine-grained knowledge of the program’s internal information, use isolated factors for seed filtering, and result in the unfairness of time consumption and gain. First, the Attributed Interprocedural Control Flow Graph (AICFG) is used to comprehensively characterize function information and functional relationships. Then, the seed is scored and evaluated in light of the characterization and then a more effective seed filtering strategy is proposed. At the same time, the attribute range of the interprocedural control flow graph is adjusted according to the number of function hits, and the graph propagation algorithm is employed to propagate attribute changes. The experimental results show that the two optimization strategies have improved the number of paths by 11.6% and 13.7% respectively compared with the baseline fuzzing tool, Azmerican Fuzzy Lop (AFL), during the testing of flvmeta. The tool FunAFL implemented also achieves higher coverage during the testing of common software such as jhead, flvmate, and libtiffin than mainstream fuzzing tools, MOPT, and FairFuzz. FunAFL finds 7 bugs and gets 1 CVE number during the test of binutils, ffjpeg, xpdf, jhead, libtiff, and libelfin.
2021, 30(11):155-163. DOI: 10.15888/j.cnki.csa.008151
Abstract:This study proposes an energy consumption prediction method based on Reinforcement learning and Generative Adversarial Networks (Re-GANs). The algorithm constructs the generator and discriminator in Generative?Adversarial?Nets (GANs) into the Agent and reward function in reinforcement learning respectively. In the training process, the current real energy consumption sequence is taken as the input state of the Agent (generator), and a set of generation sequences with a fixed length is constructed. Combined with the discriminator and Monte-Carlo search method, the reward function of the current sequence is further constructed as a reward for the first subsequent energy consumption value of the real sample sequence. On this basis, the objective function of reward is constructed, and the optimal parameters are solved. Finally, the proposed algorithm is used to predict the public building energy consumption data of the Downing Street complex. The experimental results show that the proposed algorithm has higher prediction accuracy than the multi-layer perception machine, gated loop neural network, and convolution neural network.
2021, 30(11):164-171. DOI: 10.15888/j.cnki.csa.008198
Abstract:Digital watermarking is the key technology to protect digital copyright. In this paper, the formal definition of the double watermarking algorithm is given firstly. Then, based on the Discrete Wavelet Transform (DWT), the SM4 block cipher algorithm, and the Paillier homomorphic cipher, a double watermarking algorithm in the ciphertext field is designed. When watermarking is embedded, the carrier image is transformed by triple DWT, and the band set is divided into the encryption part, the horizontal high-frequency LH3 watermarking part, and the vertical high-frequency HL3 watermarking part. The SM4 block cipher and Paillier public-key cipher are used to encrypt the frequency band coefficients of the encryption part and the watermark part, respectively. At the same time, the Paillier public-key cipher scheme is employed to encrypt the digital watermark information, and two user watermarks are embedded in the LH3 and HL3 ciphertext fields with the Least Significant Bit (LSB) method. Finally, a watermark ciphertext image is generated after the inverse wavelet transform of DWT. In the process of watermark extraction, due to the homomorphism of Paillier, the watermark can be extracted from the plaintext after decryption. Experimental results show that the algorithm is capable of fast encryption and decryption and good invisibility of watermarks.
2021, 30(11):172-178. DOI: 10.15888/j.cnki.csa.008193
Abstract:The order task allocation of autonomous mobile swarm robots in intelligent warehousing is modeled as a multi-objective optimization problem of cooperative swarm robotic scheduling, in which the path and time cost of member robots completing the picking task is viewed as the optimization objective. An ant colony-genetic algorithm fusion framework is designed. In this framework, the ant colony algorithm is taken as the secondary algorithm for initial population optimization, while the improved genetic algorithm as the main. To be specific, an elite reservation strategy is adopted after the roulette wheel selection operator in the genetic algorithm, and the inversion operator is added. A series of task allocation experiments are performed under conditions of different numbers of tasks and swarm sizes. The simulation results show that the proposed algorithm dominates over the ant colony algorithm and the genetic algorithm in performance. It combines the robustness of the ant colony algorithm and the global search ability of the genetic algorithm, improving the overall operation efficiency of the intelligent warehousing system.
2021, 30(11):179-187. DOI: 10.15888/j.cnki.csa.008179
Abstract:With the integration of geological research and big data, massive geological data with “multiple heterogeneity, high capacity and low value density” has been formed. Especially in urban construction, the monitoring data of geological deformation which reflects the ground surface and land subsidence features large capacity, time-varying property and complex dimensions. How to use visualization techniques to serve geological research analysis and problem decision more intuitively has become a hot spot of data visualization research and application in the field. To solve the problem, this paper presents a visualization method to display the monitoring situation of regional subsidence deformation under the Web three-dimensional scene constructed by the fusion of Cesium and GeoServer with the data collected by Interferometric Synthetic Aperture Radar (InSAR). Regarding the effect of deformation layer display, we create a way to dynamically change the transition color of the point cloud data from deformation monitoring, which is different from the single rendering effect of Google Earth. It is compared with ENVI/SARscape in terms of visual analysis interaction. The results of practice and application show that, in comparison with traditional methods such as two-dimensional display and integrated display of data imported from Google Earth, the proposed method can display deformation monitoring results more intuitively. In addition, it has more abundant man-machine interaction modes, which thus provides a better auxiliary decision-making function for geological professionals.
2021, 30(11):188-194. DOI: 10.15888/j.cnki.csa.008147
Abstract:In this study, we propose a forecasting model based on K-means clustering and a machine learning regression algorithm for the sales forecasting of multiple commodities in the retail industry. First, we utilize the clustering technique to identify commodities with similar sales patterns and then divide the whole dataset into different groups. Subsequently, three machine learning regression algorithms, i.e., support vector regression, random forest and XGBoost models, are trained on each sub-dataset. The data size for model training and the scope of forecasting variables are increased by the construction of a data pool. The proposed models are verified on a real sales dataset of a retail company. The experimental results show that the forecasting model based on K-means and support vector regression performs the best, and the forecasting performance of the proposed models is significantly better than that of the benchmark models and the machine learning models without using clustering.
2021, 30(11):195-202. DOI: 10.15888/j.cnki.csa.008133
Abstract:Traditional image captioning has the problems of the under-utilization of extracted image features, the lack of context information learning and too many training parameters. This study proposes an image captioning algorithm based on Vision-and-Language BERT (ViLBERT) and Bidirectional Long Short-Term Memory network (BiLSTM). The ViLBERT model is used as an encoder, which can combine image features and descriptive text information through the co-attention mechanism and output the joint feature vector of image and text. The decoder uses a BiLSTM combined with attention mechanism to generate image caption. The algorithm is trained and tested on MSCOCO2014, and the scores of evaluation criteria BLEU-4 and BLEU are 36.9 and 125.2 respectively. This indicates that the proposed algorithm is better than the image captioning based on the traditional image feature extraction combined with the attention mechanism. The comparison of generated text descriptions demonstrates that the image caption generated by this algorithm can describe the image information in more detail.
2021, 30(11):203-209. DOI: 10.15888/j.cnki.csa.008163
Abstract:Large cumulative errors of laser odometers and inaccurate rotation estimation can be encountered when the Lightweight and Ground-Optimized Lidar Odometry And Mapping (LeGO-LOAM) with line and surface feature matching is used for real-time mapping and positioning of an automated guided vehicle indoors and outdoors. In view of these problems, this work adopts the LeGO-LOAM with tightly coupled Inertial Measurement Unit (IMU) and lidar to construct the joint error function of IMU and lidar with the initial position and pose information provided by IMU for the lidar. As a result, the joint iterative optimization of position and pose is achieved. To cope with the outdoor cases with less structured information, a hybrid matching algorithm depending on the tight coupling of IMU and lidar is further proposed on the basis of the high positioning accuracy of the point-to-point Iterative Closest Point (ICP) algorithm in light of the complementarity between LeGO-LOAM and ICP algorithms. When there is much structured information in the environment, the laser odometer employs the LeGO-LOAM algorithm, and ICP algorithm functions in the case of less structured information. The experimental results show that the hybrid matching algorithm based on the tight coupling of IMU and lidar can effectively reduce the relative pose error and cumulative error of the laser odometer. In addition, it is able to eliminate some map ghosting by improving the positioning accuracy of Automated Guided Vehicles (AGVs).
2021, 30(11):210-216. DOI: 10.15888/j.cnki.csa.008153
Abstract:To improve the performance of image classification, this paper proposes an image classification algorithm based on the fusion of Multi-model Feature and Reduced Attention (MFRA). Through multi-model feature fusion, the network can learn the features of different levels of input images, increase the complementarity of features and improve the ability of feature extraction. The introduction of the attention module makes the network pay more attention to the target area and reduces the irrelevant background interference information. In this paper, the effectiveness of the algorithm is verified by a large number of experimental comparisons on three public datasets, Cifar-10, Cifar-100 and Caltech-101. The classification performance of the proposed algorithm is significantly improved as compared with existing algorithms.
2021, 30(11):217-223. DOI: 10.15888/j.cnki.csa.008171
Abstract:In the User-centric Ultra-Dense Network (UUDN), due to the limited length of the pilot sequence for channel estimation, the traditional eavesdropping detection technology based on information theoretic criteria cannot function ideally or is even completely invalid. In view of this, this study proposes a multi-node joint detection algorithm based on the LS-FDC criterion. This method uses the Linear Shrinkage (LS) theory in statistics to shrink and optimize the sample covariance matrix received by each node so that it can better fit the distribution of the overall eigenvalues after eigen-decomposition. The nodes in the Access Point Group (APG) jointly determine whether there are eavesdroppers with the Flexible Detection Criterion (FDC) algorithm. The simulation experiments and theoretical analysis show that compared with other algorithms for signal source estimation and detection, this algorithm has a significantly improved detection probability when the signal-to-noise ratio is low and the pilot sequence length is limited. A good detection effect can still be achieved even when the pilot sequence length is shorter than the number of node antennas.
2021, 30(11):224-230. DOI: 10.15888/j.cnki.csa.008192
Abstract:Abnormal driving behaviors of drivers pose a high risk of traffic accidents, threatening the life safety of drivers, passengers, and others. Thus, detecting the abnormal driving behaviors of drivers is of great significance for ensuring people’s travel safety. In an actual driving process, abnormal behaviors in the driver’s mouth region are complex and diverse. In view of this, this study proposes an unsupervised detection algorithm for the abnormal behaviors in the mouth region. The algorithm first uses the facial landmark detection network to obtain the mouth area with a high probability of anomaly. Then, the mouth area image is rebuilt with the improved Convolutional Auto-Encoder (CAE) algorithm. Abnormal behaviors are determined by the computation of the reconstruction error. The proposed algorithm is improved in three aspects: (1) the introduction of the skip connection structure to better reconstruct the input image; (2) the introduction of the Inception structure and the optimization of the proportions of branch channels to better fit the features of the input image; (3) the addition of Gaussian white noise in the training process to improve the robustness of model detection. The experimental results show that the AUC of the proposed algorithm framework is improved from 0.682 to 0.938 as compared with the traditional CAE algorithm and it can run on embedded systems.
2021, 30(11):231-239. DOI: 10.15888/j.cnki.csa.008182
Abstract:The research on the car-following model of Connected and Autonomous Vehicles (CAVs) can provide a model reference for large-scale field testing in the future, and the model has become a research hotspot in the field of traffic flow and intelligent transportation. To better study the car-following characteristics of CAVs, this study proposes a car-following model BL-MVDAM considering the multiple preceding vehicle information and backward looking effect on the basis of the MVD model. The judgment basis for the traffic flow stability of the BL-MVDAM model is deduced by linear stability analysis. The effects of different parameters in the model on the system stability are analyzed. The analysis results are verified by a simulation experiment. In the experiment, a slight disturbance is applied to a vehicle group in the car-following process on a circular road. This experiment is designed according to the attention P of a car in the group to the follower and the number k of cars in front. The speed fluctuation of the vehicle group in the proposed model is small in comparison with the FVD, MVD, OMVC and BLVD models under the same initial conditions. Especially, when P is 0.8 and k is 3, the average speed fluctuation can be as low as 0.24%. The experimental results show that the model considering the multiple preceding vehicle information and backward looking effect has a better stability region, which can better absorb the disturbance and enhance the driving stability of a vehicle group.
2021, 30(11):240-246. DOI: 10.15888/j.cnki.csa.008205
Abstract:Automatically identifying the types and counting the numbers of ships on waterways is of great significance for the construction of “smart waterways”, intelligent early warning regarding water surface, and navigation decision support. In this study, ship sample images are trained with the YOLOv3 pre-training model, and the detection model for ships on waterways is developed after parameter adjustment and optimization. Then, considering that the deep learning model is good at extracting target features, this study combines the target HSV color histogram features and LBP local features to achieve the target selection. In view of common drift and jitter of tracking targets, a correction network is designed with the integration of regression-based direction judgment and target counting with a variable time window, which realizes the automatic detection, tracking and self-correcting counting of moving targets on water surface. The test results show that the proposed method is stable and robust, suitable for automatically analyzing channel videos and extracting statistical data.
2021, 30(11):247-253. DOI: 10.15888/j.cnki.csa.008150
Abstract:To tackle the problems of the illumination interference and difficulty in capturing accurate eye center points during the human eye location by the traditional projection method, this study proposes a human eye location algorithm based on multi-scale self-quotient images with an improved integral projection method. Firstly, multi-scale self-quotient images are used to eliminate the illumination effect on the face image. Secondly, depending on the gray distribution characteristics of eyes in the horizontal direction, the integral projection method is improved with two row gradient operators to intensify eye region features and preliminarily locate the eye region. Thirdly, the eye filter image is generated after the eye region is filtered with the Sobel operator, whose vertical integral projection curve is then subjected to fitting by the Gaussian function. According to the fitting results, the left and right eye windows are segmented and, finally, their dimensions are calculated to determine the respective center point, namely the human eye center point. The tests on the YaleB face database and JAFFE face database show that the proposed method has strong adaptability to complex illumination, face edge, and face expression and can accurately locate human eye center points.
2021, 30(11):254-259. DOI: 10.15888/j.cnki.csa.008152
Abstract:Guide Positioning Sequencing (GPS) is a novel method for genome-wide DNA methylation detection. The generated sequencing data has the advantages of low detection cost and no sequence preference. At present, the most important step in methylation analysis is to align the sequences to the reference genome. However, the existing method uses Smith-Waterman for local sequence alignment, which takes too much time and affects the mapping efficiency. Therefore, a new alignment algorithm for the GPS data is proposed. The algorithm uses the advantages of paired-end sequencing to determine the alignment positions. The methylation sequences are first aligned to the reference genome, and then corresponding regular sequences are used to determine the final positions. The experimental results show that compared with the existing method, the method presented in this paper has a high mapping efficiency with comparable accuracy and the time performance improved by more than 3 times.
2021, 30(11):260-265. DOI: 10.15888/j.cnki.csa.008178
Abstract:In recent years, drones have developed rapidly in the field of logistics and transportation. An important reason is that drones could cope with various complex traffic environments such as urban traffic congestion and poor road conditions in remote rural areas. Path planning is a key part of their practical application process. This study designs an adaptive large neighborhood search algorithm for it. The algorithm improves the traditional neighborhood search by introducing an adaptive mechanism, so that it has the potential to find better solutions. Simulation experiments on some classic datasets show that the proposed algorithm has strong robustness and stability. In addition, comparative experiments with other meta-heuristic algorithms verify that the proposed algorithm can effectively reduce the cost of logistics distribution with a drone.
2021, 30(11):266-272. DOI: 10.15888/j.cnki.csa.008206
Abstract:To protect the security of carrier image content and digital watermark copyright, this study proposes a commutative encryption and watermarking algorithm integrating the Arnold permutation and histogram translation-based reversible watermarking algorithm. Taking advantage of the invariance of the image histogram after Arnold permutation, the algorithm maps the operation of watermark embedding in ciphertext to plaintext, which realizes the exchange of encryption and watermark embedding in operation sequence. The watermark can be extracted directly from the watermarked ciphertext, or be extracted after decryption. Experimental results show that the sequence of encryption and watermark embedding of this algorithm does not affect the generation of watermarked ciphertext, and the sequence of decryption and watermark extraction does not influence watermark extraction and image restoration. In addition, the direct decryption yields high-quality watermarked plaintext images, good invisibility of the watermark, and high efficiency of the algorithm.
2021, 30(11):273-280. DOI: 10.15888/j.cnki.csa.008233
Abstract:A video-based multi-object vehicle tracking and real-time trajectory distribution algorithm is proposed to display the driving trajectories of vehicles in a highway traffic video, which can provide useful traffic information for traffic management and decision-making. Firstly, the YOLOv4 algorithm is used to detect vehicle objects. Secondly, in different traffic scenarios, the vehicle data is correlated to yield a complete trajectory by using the proposed tracking method based on sparse frame detection in combination with KCF tracking algorithm. Finally, the vehicle trajectory is displayed with the vehicle distribution map and the top view of traffic scenes, which is convenient for traffic management and analysis. Experimental results show that the proposed vehicle tracking method has an excellent tracking accuracy and a fast processing speed. The real-time trajectory distribution correctly reflects the lane information of real scenes and movement information of the object vehicles, which has a great application value.
2021, 30(11):281-288. DOI: 10.15888/j.cnki.csa.008124
Abstract:To tackle the classification problem of high-dimensional group variables, this study proposes an MCP-based AdaBoost ensemble-pruning logistic regression model (AdaMCPLR). The MCP function is applied to feature selection and ensemble pruning simultaneously, which not only simplifies the model, but also effectively improves the prediction accuracy. For the efficiency enhancement, this paper improves the PICASSO algorithm to make it applicable to group variable selection. Simulation experiments show that the AdaMCPLR method is superior to other prediction methods in variable selection and classification prediction. Finally, the AdaMCPLR method proposed in this study is applied to the financial distress prediction of listed companies in China.
2021, 30(11):289-297. DOI: 10.15888/j.cnki.csa.008162
Abstract:In this work, we study the ground staff scheduling problem for airports facing a shortage of personnel. From a management perspective, the objectives are to minimize costs, address the personnel shortage problem without hiring temporary workers, and improve the responsive ability to the interruptions caused by uncertain factors. We propose a mixed-integer optimization model for balanced task coverage considering limited personnel and, according to features of the problem, design an efficient heuristic algorithm. The effectiveness of the algorithm and model is verified by real cases from a large airport. From two practical perspectives, the distribution of unassigned tasks in the scheduling cycle and the fairness among employees, we analyze the applicability of the model. The results confirm that the model performs well in confronting the staff shortage, improving the efficiency of airport operations, and assisting managers in making decisions on personnel composition.
2021, 30(11):298-303. DOI: 10.15888/j.cnki.csa.008164
Abstract:Since there is a chance to expose the user’s secret information during authentication, malicious adversaries may trace the user’s secret information and make illegal use of it, causing harm and loss of interest. For example, in the anonymous PAKA protocol based on SmartCard, there is no way to defend against the offline dictionary attack from adversaries after the SmartCard is lost. Therefore, the bilinear pairing operation, D-H difficulty and elliptic curve operation are combined with the registration and authentication, and then a new scheme is improved and designed utilizing password and smart card respectively. On the basis of the combination of a smart card with the improved password-based AKA scheme, an AKA protocol scheme relying on both the smart card and password is proposed, with the security proof given. It further improves the reliability and security of the PAKA protocol based on SmartCard and password.
2021, 30(11):304-309. DOI: 10.15888/j.cnki.csa.008165
Abstract:With the rapid development of social economy, subways, tunnels, and bridges are occupying higher positions in people’s lives. Predicting and analyzing the structural deformation data of buildings and discovering hidden safety hazards in time have become indispensable for structural safety monitoring. Combining the advantages of Long Short Time Memory (LSTM), this study proposes a structural deformation prediction model based on Bidirectional Long Short Time Memory (Bi-LSTM). The model predicts the deformation data of the current node by memorizing the rules before and after the time node and fully mines the relevant information within the deformation data. Compared with WNN, LSTM, and GRU models, this model, with RMSE, MAPE, and MAE reduced by 66.0%, 61.2%, and 66.2% respectively, proves to be an effective method for predicting structural deformation.
2021, 30(11):310-316. DOI: 10.15888/j.cnki.csa.008191
Abstract:The extraction of image feature points and descriptors is the foundation of some tasks such as SLAM, SFM, and 3D reconstruction. Preeminent algorithms for image feature point and descriptor extraction play a significant role in processing these tasks. This study accomplishes some improvements in the SuperPoint network with high robustness and good performance in the extraction of feature points and descriptors. Considering the flaws of the heavy calculation burden and massive parameters, the authors first change the ordinary convolution to depthwise separable convolution, the number of layers, and the down-sampling method. Afterward, the channel pruning algorithm is perfected so that it can be applied to depthwise separable convolution and prune the network. Experiments have proved that this study reduces the network parameter number and calculation burden respectively to 15% and 5% those of the original SuperPoint network, and the FPS is increased by 6.64 times under the condition of a slight loss of feature point detection and matching effects. Thus, good experimental results are achieved.
2021, 30(11):317-322. DOI: 10.15888/j.cnki.csa.008199
Abstract:It is difficult to measure and evaluate the line profiles of shift drums, and a single piece of system engineering software is hard to quickly realize the task requirements for line profile evaluation in real time. With the .NET object-oriented programming interface technology, the algorithm for line profile evaluation is developed by Matlab, and LabView collects and analyzes real-time data. As a result, hybrid programming is realized and software for line profile evaluation is developed. The hybrid programming technology accelerates the process from the integration of control, acquisition, analysis, and other tasks to the actual deployment and application, and improves the development efficiency.
2021, 30(11):323-328. DOI: 10.15888/j.cnki.csa.008188
Abstract:With regard to erroneous detection and low efficiency problems in the traditional manual detection of fabric defects, a self-adaptive detection method for fabric defect contours is proposed based on the visual perceptual mechanism. Firstly, the mechanism of visual information processing with the receptive field of the retina in the visual system is simulated to filter the fabric defect images and enhance the defects. Secondly, the edge information in the enhanced fabric defect images is detected according to the edge detection model of fabric defects proposed on the basis of the orientation selectivity mechanism in the area of primary visual cortex (V1). Lastly, the fabric defect contour is extracted by re-processing the detected edge images through self-adaptive threshold selection. To validate the effectiveness and accuracy of our method, this paper tests and compares four kinds of fabric defects from both qualitative and quantitative aspects. The results show that the proposed method performs well in detecting the contour information of fabric defects. This method not only acquires relatively high-quality detection images of fabric defects but also selects parameters adaptively in the whole process. It avoids the effect of subjective factors and implies practical application value.
2021, 30(11):329-335. DOI: 10.15888/j.cnki.csa.008280
Abstract:With the continuous development of the digital twin technology at this stage, research and applications surrounding digital twins have gradually become a hot spot. Because traditional autonomous driving test methods have various defects in terms of functionality, safety, and test cost, this article proposes a construction method of the digital-twin automatic driving test environment based on mixed reality in light of the basic characteristics of digital twins and the test method of autonomous driving. The autonomous driving information in the actual environment is mapped to the virtual scenario through spatial coordinate mapping, collision detection model, and virtual scene registration. In addition, the corresponding mixed reality based automatic driving test model is constructed. The collision test demonstrates that the mixed reality system has interactive features. The performance of the system at sampling frequencies of 50 ms, 200 ms, and 1000 ms is compared and analyzed. Experiments show that the algorithm in this study has better operating frame rate characteristics at the sampling frequency of 200 ms or above.
2021, 30(11):336-341. DOI: 10.15888/j.cnki.csa.008146
Abstract:To further improve the fusion effect of visual and infrared images, this paper proposes an image fusion model based on multi-scale convolution operators and DenseNet. This model first uses multi-scale convolution operators to get the direct multi-scale features of images. Then, the DenseNet is used to calculate the indirect multi-scale features of images. To get the fusion weights of image pixel information on different scales, this paper fuses the DenseNet on different scales in a stacking manner, and the fusion weights of the two kinds of images can be derived by activity graphs. At last, the fused image is derived according to the fusion weights. The experimental results show that the recognition rate is high on the THO and CMA sets.
2021, 30(11):342-347. DOI: 10.15888/j.cnki.csa.008149
Abstract:Crowd model evaluation is a key issue in virtual crowd simulation. Most studies use the error between the individual simulation trajectory and the real trajectory to evaluate the crowd model. However, crowd behavior is essentially a complex random system, and simple trajectory comparison cannot adequately reflect model capabilities. In this study, the model evaluation method based on entropy metric is used to achieve the accurate quantitative evaluation of crowd simulation by estimating the error distribution between the real crowd state and the simulated crowd state. In addition, the judgment and processing rules for distortion are introduced so that the evaluation method can be accurate under the simulated distortion. The experimental results show that the algorithm and rules proposed in this study can fully realize the quantitative evaluation of the crowd simulation model and provide guidance for the selection of model parameters.