Search

Solving the stereo camera synchronization problem or estimating the temporal offset for dynamic scenes in triangulation is crucial in computer vision and photogrammetry. However, for cheap cameras like webcams, or cameras with long distance, synchronization is costly and complex. This paper proposed a novel method to determine 3D position for dynamic scenes with unsynchronized image sequences. For the proposed method, synchronized cameras are not essential, also, the temporal offset is not essential to be known. Additionally, it can be used for linear and curve motion, which provides a wider range of applications. In linear motion, the traditional triangulation is abandoned in this paper. Two planes can be determined with unsynchronized image sequences captured by two cameras and the intersection of the two planes is the motion trajectory. Also, one plane can be determined with unsynchronized image sequences captured by one camera, one ray can be determined with 2D projection captured by another camera, and the intersection of the plane and ray represents the current position. In curve motion, the correspondence points can be determined with epipolar line and 2D trajectory determined by unsynchronized image sequences, then the current 3D position is estimated by triangulation. This means using two completely independent cameras to capture dynamic scenes, our method can achieve determination under only ‘see’ and has nothing to do with other conditions (e.g., synchronization, frame rate). Compared to several SOTA (state-of-the-art) solvers, the experimental results show our proposed method performs robustly and high precisely.

This study explores the consensus of multi-agent systems in the presence of ambient noise using intermittent Brownian noise stabilization. Firstly, the research provides a mathematical explanation of multi-agent systems and intermittent stochastic noise, and establishes a sufficient condition for multi-agent systems to achieve consensus using intermittent stochastic noise control input. Secondly, a consensus criterion is proposed for a class of multi-agent systems that are affected by ambient noise. Finally, the simulation results demonstrate that the intermittent stochastic noise stabilization technique can facilitate the establishment of consensus in multi-agent systems.

The Virtual Power Plant (VPP) serves as the primary form of participation for distributed power generation resources in electricity markets. Throughout this market engagement, uncertainties stemming from renewable energy sources and market prices present significant revenue risks for VPPs. Addressing this challenge, the paper proposes a VPP bidding risk management model centered on dominance constraints. In contrast to conventional methods reliant on risk metrics like Conditional Value at Risk, the proposed model offers improved control over downside and tail risks while reducing optimization iterations to just one, thus substantially reducing computational cost associated with risk management. The study also contrasts the efficacy and computational costs of first-order and second-order dominance models. For instance, in a three-scenario benchmark, the first-order dominance model exhibits a 16.7% lower overall risk level compared to the second-order model. Similarly, in a nine-scenario benchmark, the computational time of the first-order model increases by 0.96% due to the introduction of binary variables, and this increase scales with the number of benchmark scenarios. The paper additionally incorporates mean deviation of profits and expected excess returns to develop a bi-objective optimization model. This model not only fulfills risk control requisites but also shapes return distributions to align more closely with VPP decision-makers’ profit preferences. The case study is based on a VPP comprising conventional thermal units, wind turbines, energy storage stations, and flexible loads.

With the increasing demands on equipment accuracy, synchronous performance becomes more and more significant for ultra-precision systems. In order to achieve nanometer-level accuracy and millisecond synchronization in practice, a learning synchronous control (LSC) strategy has been proposed in this research. Specifically, in our research, the LSC strategy includes Genetic Algorithm (GA) tuning method and Back Propagation (BP) neural network synchronous controller. Herein, GA tuning method has been designed to optimize the parameters of controllers for each axis offline. In order to enhance real-time information exchange between axes, a BP synchronous controller has been designed to enhance the coupling between axes and compensate for dynamic disturbances. Then the convergence criteria and stability analysis of the method have been presented. The proposed LSC was simulated and experimented on an ultra-precision stage system, and the HIL experimental results demonstrate that the proposed LSC can simultaneously possess nanometer-level tracking/synchronous accuracy and millisecond synchronization. Furthermore, compared to the conventional proportion integration differentiation (PID) method, the proposed LSC has higher synchronization performance with MA of 0.64 nm, MSD of 3.98 nm, and a settling time of 10 ms. The LSC strategy provides an effective control technology with good potential in industrial applications.

Recently, the combination of convolutional neural network (CNN) and long short-term memory (LSTM) exhibits better performance than single network architecture. Most of these studies connect LSTM networks behind CNNs. When operating on hardware, the current design of CNN-LSTM is similar to a pipeline architecture. However, the classic structure lead to a feature loss when data is sent to LSTM since CNN is not good at extracting temporal features. At the same time, as the depth and scale increases, it will bring a huge amount of computation, which makes hardware implementation difficult. Based on that, a parallel CNN-LSTM architecture is proposed, in which two networks extract features from the input data synchronously, being proven to be more effective than classical CNN-LSTM. This paper designs a parallel CNN-LSTM computing device based on FPGA. The device is divided into control unit and operation unit. Control stream and data stream transport between the two units, ensuring the proper running of the device. A highly parallel multi-channel convolution layer and pooling layer are designed to improve the calculation efficiency. A 4-stage pipeline structure is adopted to implement the LSTM part. This paper makes full use of on-chip BRAM to design a look-up table for activation function approximation, reducing the resource consumption by 95% compared with the traditional polynomial approximation. Finally, we verify our device under cooperative spectrum sensing (CSS) and handwritten classification scenarios. Proposed device reaches higher accuracy in two scenarios compared with classic CNN-LSTM structure as well as faster calculating speed, and the overall project power is limited below 2W. The scalability and limitation of this computing device are also discussed.

Cognitive radio (CR), as an emerging technology to improve the utilization of radio spectrum, the fundamental of CR technology is spectrum sensing, due to the detection performance being affected by various factors, spectrum sensing is challenging to achieve accurately. In recent years, many spectrum sensing algorithms have been proposed, such as energy detection algorithm, matched filter detection algorithm, cyclic stationary detection algorithm, etc. However, these algorithms are model-driven and require certain prior information. If the model assumptions are inaccurate or the prior information is challenging to obtain, the algorithms’ detection performance will be degraded. The development of artificial intelligence technology and deep learning provides a new way to realize spectrum sensing. In this paper, we design a cooperative spectrum sensing model based on the parallel connection of convolutional neural network (CNN) and long-short-term memory (LSTM), which makes full use of the complementary feature extraction capabilities of CNN and LSTM networks. Among them, CNN is used to extract hidden spatial features, and LSTM network is used to extract time features. Both CNN and LSTM can process the original dataset directly avoiding information feature loss when the network is connected serially. Experimental result shows that the detection performance of the proposed algorithm outperforms the conventional cooperative detection algorithm under low SNR condition. For example, when the number of cooperative users is 9 and the transmit power is 10, the detection probability of the proposed algorithm in this paper can reach more than 90%, which is much higher than the detection performance of other spectrum detection algorithms.

2D-3D point or line correspondences are usually used to estimate the camera pose and intrinsic parameters, such as the focal length. There are many point-based methods and line-based methods, i.e., PnP solvers and PnL solvers. However, methods that use both point and line correspondences are not used widely. Hence, this paper proposed a new method to estimate the pose and focal length using a 2D-3D point correspondence and a 2D-3D line correspondence. In addition, the camera position is known as the prior knowledge to simplify the calculating process. In the world frame, the 3D point and camera position can determine a line, and the 2D projection of the 3D point and camera position can also determine a line in the camera frame. In fact, the two lines are the same line in space and their unit direction vectors can convert to each other using the rotation matrix from the world frame to the camera frame. In the world frame, the 3D line and camera position can determine a plane, and the 2D projection of the 3D line and camera position can also determine a plane in the camera frame. In fact, the two planes are the same plane in space and their unit normal vectors can also convert to each other using the rotation matrix from the world frame to the camera frame. Now there are two transformations of the direction vectors and the normal vectors. The transformations of vectors can be regarded as the transformations of 3D-3D point correspondences when the world frame and camera frame have the same origin. This is the key point in this paper and then the pose can be estimated efficiently. What’s more, the angle between the line passing through the 3D point and camera position, and the plane passing through the 3D line and camera position in the world frame, is equal to the angle between the line passing through the 2D projection of the 3D point and camera position, and the plane passing through the 2D projection of the 3D line and camera position in the camera frame. This is another key point to estimate the focal length. Experiments show our proposed method has good performance in numerical stability, noise sensitivity and computational speed with synthetic data and real scenarios. In addition, compared with other SOTA (state-of-the-art) methods, our method has a single solution and performs better.

Indoor localization with high accuracy plays a key role in the field of Internet of Things in 5th-Generation (5G) era. With the introduction of Multiple Input Multiple Output (MIMO) technology in 5G, the direction-of-arrival (DOA) method is highly feasible in indoor localization. However, the direction of arrival is susceptible to complex indoor environment. To improve the accuracy and stability of DOA estimation, an adjacent angle power difference (AAPD) method is proposed based on Orthogonal Matching Pursuit (OMP). This method uses OMP to obtain an initial estimate of the direction, and then, adjusts the estimation by calculating the difference power of adjacent points at initial value point to get the fractional DOA. In the scenario of continuous movement, beamforming is further applied, which reduces the amount of calculation. Both simulation and experimental results show that the proposed method can achieve high accuracy and eliminate the error jitter. Compared with the classical Multiple Signal Classification (MUSIC) method for DOA estimation, the proposed method can increase accuracy by 46% under the condition of low SNR (Signal-to-Noise Ratio). The probability that the measurement error does not exceed 5° in the actual movement tests is 97.5%.

This paper proposes a sub-frame layer RDO algorithm to further improve coding performance of low-delay hierarchical prediction structure in High Efficiency Video Coding. In the proposed algorithm, each frame is segmented into two sub-frames according to motion intensity, and a sub-frame layer is taken as a basic optimizing unit. An inter-frame dependency model is proposed to quantify inter-frame dependency of a frame or a sub-frame. The proposed inter-frame dependency model is composed of two parts, Laplace distribution based direct inter-frame dependency model and distortion propagation chain. The direct inter-frame dependency model is used to evaluate dependency caused by direct reference, and the distortion propagation chain shows the relationship between direct reference and indirect reference. With quantified inter-frame dependency, Lagrange optimization is then used to solve the optimal $\lambda $ -QP pair for each sub-frame layer in frame layer 1 (L1) to frame layer 4 (L4). For frame layer 0 (L0), additional L0 frames are adaptively inserted according to inter-frame dependency of the previous L0 frame, and the optimal $\lambda $ -QP pair for each sub-frame in L0 is solved using Lagrange optimization. When peak signal to noise ratio is used as quality criterion, the proposed algorithm achieves 7.0% and 7.0% bit-rate reduction with lowdelay_P_main (LDP) and lowdelay_main (LDB) configurations respectively; when structural similarity is used as quality criterion, the proposed algorithm achieves 13.3% and 13.1% bit-rate reduction with LDP and LDB configurations respectively.

Systematic operations make modern wars more and more complicated, thus it is a challenge to design, develop, and select an excellent System-of-Systems. In this paper, we propose an equipment System-of-Systems architecture (ESoSA) formalization framework based on the super network and an architecture space searching algorithm. First, the formal definition of ESoSA and the problem of ESoSA space searching problem framework are introduced, where the searching problem is transferred into a multi-agent dynamic programming problem. Each agent searches for the optimal architecture in its corresponding architecture space, and different architecture spaces may overlap with each other. Second, the objective of the problem is to choose several optimal architectures with maximized sampling rewards and minimized cumulative searching costs. Third, a sequential search algorithm based on decision indicators is proposed to solve the problem. Finally, the proposed algorithm is evaluated by simulation experiments, where the experimental results show that the proposed algorithm is with high quality, and outperforms other benchmark algorithms.

In this paper, an adaptive channel estimation and tracking method is proposed for the uniform rectangular array (URA) based massive multiple-input and multiple-output (M-MIMO) systems. To decrease the computational complexity while tracking the change on the channel state information (CSI) in time, a principal subspace analysis (PSA) based scheme is proposed to estimate and track the signal subspace based on the received signal covariance matrix. Based on the estimated signal subspace, the channel parameters including the direction of arrival (DoA) angle and the channel gain corresponding to each resolvable path of the channel, can be recovered via the classic multiple signal classification (MUSIC) algorithm. Then, the Cramer-Rao lower bound (CRLB) is derived to evaluate the performance of the proposed scheme. Simulation results are provided to verify the accuracy of the proposed scheme on the channel estimation.

One major reason for the success of convolutional neural networks (CNNs) is the availability of large-scale labeled data. Effective training of CNNs relies on large annotated data. Unfortunately, large amounts of data with corresponding annotations are too expensive to obtain in some real-world applications. One reasonable alternative is to use data augmentation techniques to automatically generate annotated samples. In this paper, a novel data augmentation framework based on perspective transformation is proposed. This method automatically generates new annotated data without extra manual labeling, thus effectively extends the inadequate dataset. Perspective transformation can produce new images captured from any cameras viewpoints. Therefore, our method can mimic images taken at the angle that the camera cannot reach. Extensive experimental results on several datasets have demonstrated that our perspective transformation data augmentation strategy is an effective tool when using deep CNNs on small or imbalance datasets.

Tissue equivalent liquid plays an important role in specific absorption rate (SAR) assessment. Proficiency tests (PTs) through interlaboratory comparison of dielectric parameters (i.e., dielectric constant and the electrical conductivity of the tissue equivalent liquid) were conducted in 2017 and 2019. We coordinated the PTs performed with traveling samples circulating among the participating laboratories. As one of the most crucial issues in the PT schemes, these samples are described comprehensively, including the realization, the assignment of reference values, and the verification of homogeneity and stability. Fifteen and eleven laboratories participated in the PTs in 2017 and 2019, respectively. In total, 74 measurement results were provided in terms of best estimate and uncertainty. This paper presents and evaluates the performance of all the participating laboratories. Additionally, the measurement results are grouped in terms of the types of measurement probes. The analysis indicates that the measurement results obtained from two most widely used types of probes are compatible.

Occupants' comfort level has a strong correlation with health problems. Providing a comfortable environment for the occupants will bring the benefits of improved health. To achieve this goal, it is necessary to have a reliable human comfort model for predicting the occupants' comfort level and subsequently controlling the involved comfort condition. However, the comfort perception of occupants is subjective. There is a lack of objective indices for measuring comfort level. Furthermore, human comfort is affected by various environmental factors. Such situations make it difficult to set up a model for measuring human comfort. To address the challenges, we use Blood Pulse Wave (BPW) as an objective comfort index and adopt a data-driven approach to predict human comfort level based on data including both environmental factors and human factors. We propose a framework for collecting the data followed by investigating the relationship between the factors with the purpose of building a scalable comfort model. In consideration of the nonlinear relationship present in the dataset, we opt for support vector regression with radial basis function (SVR-RBF) algorithm to establish the comfort model. To validate the predication performance of this method, we have applied the other six popular machine learning models on the same dataset. In order to choose an optimal model, we apply the holdout method and k-folder cross-validation method together with the grid search. The comparison results show that the SVR-RBF has the best performance for comfort prediction according to the mean squared error, mean absolute error and R-squared score.

Information extraction is a crucial technology to construct a knowledge base. In this paper, a novel model was proposed to extract entities and relations from plain text. This model consists of two components: a joint network and a reinforcement learning agent. The joint network is designed end-to-end, which can extract entities and relations simultaneously. In the joint model, a new tagging scheme was adopted, then the entity and relation extraction can be modeled as a joint sequence tagging problem. To enhance the robustness of the model, we also introduced a reinforcement learning (RL) agent to remove the noisy data from the training dataset. The RL agent aims at determining whether a candidate instance should be removed from the training dataset or not. When the agent completes a selection process, the training dataset will be divided into two parts: clean data and noisy data. Then the joint network can be trained again on the clean dataset to generate a better model. To assess the validity of the model we proposed, extensive experiments were conducted on the New York Times dataset (NYT10 and NYT 11). The experimental results showed that the model we proposed is superior compared with the baselines, achieving the F1 value on NYT10 and NYT11 with 0.612 and 0.549, respectively.

The systematic and network-centric warfare becomes complicated currently. It is a challenge problem to design and choose the optimal combat system-of-systems to complete missions. In order to solve this challenge, we propose a super-network model based combat system-of-systems architecture (CSoSA) and a scheme space exploration algorithm in this paper. First, a formal definition of CSoSA is given based on the operational capability generation elements. The framework for CSoSA scheme space exploration problem is then presented. The goal of the problem is to select the optimal architecture with maximized expected rewards and minimized cumulative costs. In addition, a novel and optimal algorithm for CSoSA dynamic exploration is proposed based on the greedy search algorithm. Finally, we evaluated our algorithm through simulation experiments, where the results showed that our algorithm is significantly better than several benchmark algorithms. In general, the proposed architecture framework and the exploration algorithm can assist commanders in making further decisions.

The explosive growth of massive data generation from Internet of Things in industrial, agricultural and scientific communities has led to a rapid increase for data analytics in cloud data centers. The ubiquitous and pervasive demand for near-data processing urges the edge computing paradigm in recent years. Edge computing is promising for less network backbone bandwidth usage and thus less data center side processing pressure, as well as enhanced service responsiveness and data privacy protection. Computation offloading plays a crucial role in edge computing in terms of network packets transmission and system responsiveness through dynamic task partitioning between cloud data centers and edge servers and edge devices. In this paper a thorough literature review is conducted to reveal the state-of-the-art of computation offloading in edge computing. Various aspects of computation offloading, including energy consumption minimization, Quality of Services guarantee, and Quality of Experiences enhancement are surveyed. Moreover, resource scheduling approaches, gaming and tradeoffing among system performance and overheads for computation offloading decision making are also reviewed.

In this paper, a new optimal multi-agent continuous patrol algorithm is proposed to solve the information gathering problem in dynamic environments. First, the environment is modeled as a layout graph with information attached to vertices. Each agent patrols within a specified area and only interacts with its adjacent agents. The problem is then cast as the factored multi-agent partially observable Markov decision process (MPOMDP). Furthermore, a scalable centralized online planning algorithm, called the factored belief-based variable eliminated Monte Carlo planning algorithm, is proposed based on the Monte Carlo tree search (MCTS) method. The proposed algorithm constructs an independent local look-ahead tree for each agent, where actions are coordinated at specific locations of each tree based on the variable elimination algorithm. Finally, we mimic typical patrol problems to empirically evaluate the proposed algorithm by benchmarking it against some state-of-the-art solvers. The results demonstrate that the performance of the proposed algorithm is remarkable for multi-agent systems with the weakly-coupled structure in partially observable scenarios.

Cyber physical systems (CPSs) are rapidly developing, with increasing scale, complexity and heterogeneity. However, testing CPSs systematically to ensure that they operate with high reliability remains a big challenge. Therefore, it is necessary to summarize existing works and technologies systematically, with the aim of inspiring new inventions for more efficient CPS testing. Accordingly, this paper first investigated the advances in CPS testing methods from ten aspects, including different testing paradigms, technologies, and some non-functional testing methods (including security testing, robust testing, and fragility testing). Then, we further elaborate on the infrastructures of CPS testbeds from the perspectives of their architecture and the corresponding function analyses. Finally, challenges and future research directions are identified and discussed. It can be concluded that future CPS testing should focus more on the combination of different paradigms and technologies for multi-objective by integrating more emerging cutting-edge technologies such as Internet of things, big data, cloud computing and AI.

In this paper, three easily implemented hardware algorithms, including the adaptive prediction error filter based on the Gram-Schmidt algorithm (GS-APEF), the least mean square adaptive filter and the comb filter, are extensively investigated for artifact denoising on a constructed semi-simulated database with varied ten-fold frequency stimulation. By implementing the GS-APEF in the field-programmable gate array (FPGA) and using the edge noise mitigating technique, a stimulation artifact denoising system is designed to realize real-time stimulation artifact removal under varied ten-fold frequency functional electrical stimulation. Good performance of the artifact denoising is demonstrated in proof-of-concept experiments on able-bodied subjects with a mean correlation coefficient between the root mean square profile of denoised surface electromyography and volitional force of 0.94, verifying the validity of the proposed prototype.

Most CNN-based super-resolution networks require a large number of samples for model training, which may cause overfitting when trained on a specific set, and the internal self-similarity of the test image in this way has also been discarded. To resolve this problem, this paper proposes a novel zero-shot learning (ZSL) network for super resolution. In this model, the training samples are cropped from the input image, which means no extra dataset is required for model training, the overfitting problem in this way can be well avoided. Besides, since all the training samples are cropped from the input itself, the nonlocal self-similarity attributes of the test image can be fully utilized. Moreover, the efficient spatial pyramid of dilated convolutions network with bottleneck (ESP-BNet) is applied in the model as an efficient computational structure to enhance the feature representation. Comparison experiments show that the proposed approach achieves PSNR/SSIM of 31.13dB/87.9% on Set5 dataset and PSNR/SSIM of 27.62dB/78.1% on Set14 dataset, which is around 3dB better than the traditional method and about 1dB higher than those of the state-of-art method.