As a cutting-edge branch of unmanned aerial vehicle (UAV) technology, the cooperation of a group of UAVs has attracted increasing attention from both civil and military sectors, due to its remarkable merits in functionality and flexibility for accomplishing complex extensive tasks, e.g., search and rescue, fire-fighting, reconnaissance, and surveillance. Cooperative path planning (CPP) is a key problem for a UAV group in executing tasks collectively. In this paper, an attempt is made to perform a comprehensive review of the research on CPP for UAV groups. First, a generalized optimization framework of CPP problems is proposed from the viewpoint of three key elements, i.e., task, UAV group, and environment, as a basis for a comprehensive classification of different types of CPP problems. By following the proposed framework, a taxonomy for the classification of existing CPP problems is proposed to describe different kinds of CPPs in a unified way. Then, a review and a statistical analysis are presented based on the taxonomy, emphasizing the coordinative elements in the existing CPP research. In addition, a collection of challenging CPP problems are provided to highlight future research directions.
We present a real-time monocular simultaneous localization and mapping (SLAM) system with a new distributed structure for multi-UAV collaboration tasks. The system is different from other general SLAM systems in two aspects: First, it does not aim to build a global map, but to estimate the latest relative position between nearby vehicles; Second, there is no centralized structure in the proposed system, and each vehicle owns an individual metric map and an ego-motion estimator to obtain the relative position between its own map and the neighboring vehicles’. To realize the above characteristics in real time, we demonstrate an innovative feature description and matching algorithm to avoid catastrophic expansion of feature point matching workload due to the increased number of UAVs. Based on the hash and principal component analysis, the matching time complexity of this algorithm can be reduced from O(log N) to O(1). To evaluate the performance, the algorithm is verified on the acknowledged multi-view stereo benchmark dataset, and excellent results are obtained. Finally, through the simulation and real flight experiments, this improved SLAM system with the proposed algorithm is validated.
We study a mobile edge computing system assisted by multiple unmanned aerial vehicles (UAVs), where the UAVs act as edge servers to provide computing services for Internet of Things devices. Our goal is to minimize the energy consumption of this system by planning the trajectories of UAVs. This problem is difficult to address because when planning the trajectories, we need to consider not only the order of stop points (SPs), but also their deployment (including the number and locations) and the association between UAVs and SPs. To tackle this problem, we present an energy-efficient trajectory planning algorithm (TPA) which comprises three phases. In the first phase, a differential evolution algorithm with a variable population size is adopted to update the number and locations of SPs at the same time. In the second phase, the k-means clustering algorithm is employed to group the given SPs into a set of clusters, where the number of clusters is equal to that of UAVs and each cluster contains all SPs visited by the same UAV. In the third phase, to quickly generate the trajectories of UAVs, we propose a low-complexity greedy method to construct the order of SPs in each cluster. Compared with other algorithms, the effectiveness of TPA is verified on a set of instances at different scales.
Deep reinforcement learning (RL) has become one of the most popular topics in artificial intelligence research. It has been widely used in various fields, such as end-to-end control, robotic control, recommendation systems, and natural language dialogue systems. In this survey, we systematically categorize the deep RL algorithms and applications, and provide a detailed review over existing deep RL algorithms by dividing them into modelbased methods, model-free methods, and advanced RL methods. We thoroughly analyze the advances including exploration, inverse RL, and transfer RL. Finally, we outline the current representative applications, and analyze four open problems for future research.
In underwater scenes, the quality of the video and image acquired by the underwater imaging system suffers from severe degradation, influencing target detection and recognition. Thus, restoring real scenes from blurred videos and images is of great significance. Owing to the light absorption and scattering by suspended particles, the images acquired often have poor visibility, including color shift, low contrast, noise, and blurring issues. This paper aims to classify and compare some of the significant technologies in underwater image defogging, presenting a comprehensive picture of the current research landscape for researchers. First we analyze the reasons for degradation of underwater images and the underwater optical imaging model. Then we classify the underwater image defogging technologies into three categories, including image restoration approaches, image enhancement approaches, and deep learning approaches. Afterward, we present the objective evaluation metrics and analyze the state-of-the-art approaches. Finally, we summarize the shortcomings of the defogging approaches for underwater images and propose seven research directions.
Although deep neural networks (DNNs) have achieved great success in semantic segmentation tasks, it is still challenging for real-time applications. A large number of feature channels, parameters, and floating-point operations make the network sluggish and computationally heavy, which is not desirable for real-time tasks such as robotics and autonomous driving. Most approaches, however, usually sacrifice spatial resolution to achieve inference speed in real time, resulting in poor performance. In this paper, we propose a light-weight stage-pooling semantic segmentation network (SPSSN), which can efficiently reuse the paramount features from early layers at multiple stages, at different spatial resolutions. SPSSN takes input of full resolution 2048 × 1024 pixels, uses only 1.42 × 106 parameters, yields 69.4% mIoU accuracy without pre-training, and obtains an inference speed of 59 frames/s on the Cityscapes dataset. SPSSN can run directly on mobile devices in real time, due to its light-weight architecture. To demonstrate the effectiveness of the proposed network, we compare our results with those of state-of-the-art networks.
We propose a new asymmetric pixel confusion algorithm for images based on the Rivest-Shamir-Adleman (RSA) public-key cryptosystem and Arnold map. First, the RSA asymmetric algorithm is used to generate two groups of Arnold transform parameters to address the problem of symmetrical distribution of Arnold map parameters. Second, the image is divided into blocks, and the first group of parameters is used to perform Arnold confusion on each sub-block. Then, the second group of parameters is used to perform Arnold confusion on the entire image. The image correlation is thereby fully weakened, and the image confusion degree and effect are further enhanced. The experimental results show that the proposed image pixel confusion algorithm has better confusion effect than the classical Arnold map based confusion and the row-column exchange based confusion. Specifically, the values of gray difference are close to one. In addition, the security of the new confusion operation is dependent on RSA, and it can act as one part of a confusion-substitution structure in a cipher.
Success has been obtained using a semi-supervised graph analysis method based on a graph convolutional network (GCN). However, GCN ignores some local information at each node in the graph, so that data preprocessing is incomplete and the model generated is not accurate enough. Thus, in the case of numerous unsupervised models based on graph embedding technology, local node information is important. In this paper, we apply a local analysis method based on the similar neighbor hypothesis to a GCN, and propose a local density definition; we call this method LDGCN. The LDGCN algorithm processes the input data of GCN in two methods, i.e., the unbalanced and balanced methods. Thus, the optimized input data contains detailed local node information, and then the model generated is accurate after training. We also introduce the implementation of the LDGCN algorithm through the principle of GCN, and use three mainstream datasets to verify the effectiveness of the LDGCN algorithm (i.e., the Cora, Citeseer, and Pubmed datasets). Finally, we compare the performances of several mainstream graph analysis algorithms with that of the LDGCN algorithm. Experimental results show that the LDGCN algorithm has better performance in node classification tasks.
In traditional target tracking methods, the angle error and range error are often measured by the empirical value, while observation noise is a constant. In this paper, the angle error and range error are analyzed. They are influenced by the signalto-noise ratio (SNR). Therefore, a model related to SNR has been established, in which the SNR information is applied for target tracking. Combined with an advanced nonlinear filter method, the extended Kalman filter method based on the SNR model (SNR-EKF) and the unscented Kalman filter method based on the SNR model (SNR-UKF) are proposed. There is little difference between the SNR-EKF and SNR-UKF methods in position precision, but the SNR-EKF method has advantages in computation time and the SNR-UKF method has advantages in velocity precision. Simulation results show that target tracking methods based on the SNR model can greatly improve the tracking performance compared with traditional tracking methods. The target tracking accuracy and convergence speed of the proposed methods have significant improvements.