The Journal of Supercomputing最新文献

At present, the top-down approach requires the introduction of pedestrian detection algorithms in multi-person pose estimation. In this paper, we propose an end-to-end trainable human pose estimation network named Ultra-FastNet, which has three main components: shape knowledge extractor, corner prediction module, and human body geometric knowledge encoder. Firstly, the shape knowledge extractor is built using the ultralightweight bottleneck module, which effectively reduces network parameters and effectively learns high-resolution local representations of keypoints; the global attention module was introduced to build an ultralightweight bottleneck block to capture keypoint shape knowledge and build high-resolution features. Secondly, the human body geometric knowledge encoder, which is made up of Transformer, was introduced to modeling and discovering body geometric knowledge in data. The network uses both shape knowledge and body geometric knowledge which is called knowledge-enhanced, to deduce keypoints. Finally, the pedestrian detection task is modeled as a keypoint detection task using the corner prediction module. As a result, an end-to-end multitask network can be created without the requirement to include pedestrian detection algorithms in order to execute multi-person pose estimation. In the experiments, we show that Ultra-FastNet can achieve high accuracy on the COCO2017 and MPII datasets. Furthermore, experiments show that our method outperforms the mainstream lightweight network.

Currently, Raft, as an mainstream consensus mechanism, has received widespread attention. Partition consensus can reduce the number of nodes involved in a single consensus and improve consensus efficiency. However, existing algorithms suffer from unreasonable partitioning and intolerance of Byzantine nodes. To address these problems, this paper proposes a novel Raft consensus algorithm combining comprehensive evaluation partitioning and Byzantine fault tolerance, CB-Raft. First, a comprehensive evaluation of nodes is conducted from the perspectives of consensus behavior and location, and the nodes are evenly divided based on the parity of the comprehensive ranking. Second, the leader is selected from the nodes with the top rankings in the comprehensive evaluation, and the nodes communicate with each other based on BLS signatures. Finally, a fast response mechanism based on cross-partition leader-follower communication is proposed to avoid the continued evil behavior of the leader, and a pipeline mechanism based on changeable signature thresholds is proposed to solve consensus blocking. The experimental results show that compared with the existing partitioning methods, the proposed partitioning scheme has significant advantages in terms of consensus latency, throughput, and the probability of partition success. Compared with the similar Raft algorithms, CB-Raft has high consensus performance and good resistance to Byzantine nodes.

Fully homomorphic encryption (FHE) is capable of handling sensitive encrypted data in untrusted computing environments. The efficient application of FHE schemes in secure outsourced computation can effectively address security and privacy concerns. This paper presents a novel fully homomorphic encryption scheme called GMS, based on the n-secret learning with errors (LWE) assumption. By utilizing block matrix and decomposition technology, GMS achieves shorter encryption and decryption times and smaller ciphertext sizes compared to existing FHE schemes. For secure outsourced matrix multiplication ({textbf {A}}_{mtimes n}cdot {textbf {B}}_{ntimes l}) with arbitrary dimensions, GMS only requires (O(max {m,n,l})) rotations and one homomorphic multiplication. Compared to the state-of-the-art methods, our approach stands out by achieving a significant reduction in the number of rotations by a factor of (O(log max {n, l})), along with a decrease in the number of homomorphic multiplications by a factor of n and (O(min {m, n, l})). The experimental results demonstrate that GMS shows superior performance for secure outsourced matrix multiplication of any dimension. For example, when encrypting a (64times 64)-dimensional matrix, the size of the ciphertext is only 1.27 MB. The encryption and decryption process takes approximately 0.2 s. For matrix multiplication ({textbf {A}}_{64times 64}cdot {textbf {B}}_{64times 64}), the runtime of our method is 39.98 s, achieving a speedup of up to 5X and 2X.

The typical procedure for Earth Observation Nanosatellites involves the sequential steps of image capture, onboard storage, and subsequent transmission to the ground station. This approach places significant demands on onboard resources and encounters bandwidth limitations; moreover, the captured images may be obstructed by cloud cover. Many current deep-learning methods have achieved reasonable accuracy in cloud detection. However, the constraints posed by nanosatellites specifically in terms of memory and energy present challenges for effective onboard Artificial Intelligence implementation. Hence, we propose an optimized tiny Machine learning model based on the U-Net architecture, implemented on STM32H7 microcontroller for real-time cloud coverage prediction. The optimized U-Net architecture on the embedded device introduces Depthwise Separable Convolution for efficient feature extraction, reducing computational complexity. By utilizing this method, coupled with encoder and decoder blocks, the model optimizes cloud detection for nanosatellites, showcasing a significant advancement in resource-efficient onboard processing. This approach aims to enhance the university nanosatellite mission, equipped with an RGB Gecko imager camera. The model is trained on Sentinel 2 satellite images due to the unavailability of a large dataset for the payload imager and is subsequently evaluated on gecko images, demonstrating the generalizability of our approach. The outcome of our optimization approach is a 21% reduction in network parameters compared to the original configuration and maintaining an accuracy of 89%. This reduction enables the system to allocate only 61.89 KB in flash memory effectively, resulting in improvements in memory usage and computational efficiency.

The Internet of Things (IoT) is evolving rapidly, increasing demand for safeguarding data against routing attacks. While achieving complete security for RPL protocols remains an ongoing challenge, this paper introduces an innovative hybrid autoencoder–decision tree framework (HADTF) designed to detect four types of RPL attacks: decreased rank, version number, DIS flooding, and blackhole attacks. The HADTF comprises three key components: enhanced feature extraction, feature selection, and a hybrid autoencoder–decision tree classifier. The enhanced feature extraction module identifies the most pertinent features from the raw data collected, while the feature selection component carefully curates’ optimal features to reduce dimensionality. The hybrid autoencoder–decision tree classifier synergizes the strengths of both techniques, resulting in high accuracy and detection rates while effectively minimizing false positives and false negatives. To assess the effectiveness of the HADTF, we conducted evaluations using a self-generated dataset. The results demonstrate impressive performance with an accuracy of 97.41%, precision of 97%, recall of 97%, and F1-score of 97%. These findings underscore the potential of the HADTF as a promising solution for detecting RPL attacks within IoT networks.

Wireless Sensor Networks have become an integral part of our lives with the advancement in the field of Internet of Technology. Multiple sensors operate together in Wireless Sensor Networks (WSNs) to collect data and communicate wirelessly with one another. For each sensor node’s data collection to be useful, it is essential to explore precise localization technology for WSNs. DV-Hop, as an easily implementable range-free localization algorithm, has gained significant popularity in the research community. As a result, many enhanced DV-Hop variations have been put out in the literature. However, the challenges of poor location accuracy associated with DV-Hop continue to spark interest among researchers, leading to further investigations and making it a preferred area for research in localization algorithms. Research in this paper proposes an improved version of three-dimensional DV-Hop algorithm based on improved adaptive differential evolution (3D-IADE DV-Hop). The proposed method optimizes the estimated coordinates using an improved version of adaptive differential evolution by controlling offspring generation behaviour. Moreover, we have demonstrated the superiority of 3D-IADE DV-Hop compared to other algorithms under consideration. The simulation results serve to strengthen our observations, confirming that the proposed algorithm outperforms its counterparts with enhanced performance and superiority.

With the continuous development of socio-economics, the rapid increase in the use of road vehicles has led to increasingly severe issues regarding traffic accidents. Timely and accurate detection of road traffic accidents is crucial for mitigating casualties and alleviating traffic congestion. Consequently, road traffic accident detection has become a focal point of research recently. With the assistance of advanced technologies such as deep learning, researchers have designed more accurate and effective methods for detecting road traffic accidents. However, deep learning models are often constrained by the scale and distribution of their training datasets. Presently, datasets specifically tailored for road traffic accident detection suffer from limitations in scale and diversity. Furthermore, influenced by the recent surge in research on intelligent driver assistance systems, datasets from the surveillance perspective (the third-person viewpoint) are fewer than those from the driver’s perspective (the first-person viewpoint). Considering these shortcomings, this paper proposes a new dataset, Traffic Accident Detection from the Perspective of Surveillance (TADS). To the best of our knowledge, we are the first to attempt to detect traffic accident under the surveillance perspective with the aid of eye-gaze data. Leveraging the special data components within this dataset, we design the RF-RG model (input: the RGB and optical flow values of the frames; output: the RGB and gaze values of the predicted frame) for detecting road traffic accidents from a surveillance perspective. Comparative experiments and analyses are conducted with existing major detection methods to validate the efficacy of the proposed dataset and the approach. The TADS dataset has been made available at: https://github.com/cyc-gh/TADS/.

Internet of Medical Things (IoMT) has gradually become the main solution for smart healthcare, and cloud-assisted IoMT is becoming a critical computing paradigm to achieve data collection, fine-grained data analysis, and sharing in healthcare domains. Since IoMT data can be frequently shared for accurate diagnosis, prognosis prediction, and health counseling, how to solve the contradiction between data sharing and privacy protection for IoMT data is a challenge problem. Besides, the cloud-assisted medical system is still at risk of a single point of failure and usually suffers from poor scalability and large response delay. Hence, we propose a blockchain-based privacy-preserving and secure sharing scheme for IoMT data, named BAPS. In BAPS, the Interplanetary File System (IPFS) is adopted to store encrypted records. Then, a non-interactive zero-knowledge proof protocol is employed to verify whether the stored data meets the specific request from data requesters without disclosing personal privacy. Moreover, we combine cryptographic primitives and decentralized smart contracts to achieve user anonymity. Finally, we leverage blockchain and proxy re-encryption to achieve fine-grained sharing of healthcare data. Security analysis indicates that this scheme meets the expected security requirements. The computational cost of BAPS is reduced by about 6% compared to state-of-the-art schemes, while the communication overhead is reduced by about 8%. Both theoretical analysis and experiment results show that this scheme can realize privacy-preserving and secure data sharing with acceptable computational and communication costs.

In recent years, some researchers have combined deep learning methods such as semantic segmentation with a visual SLAM to improve the performance of classical visual SLAM. However, the above method introduces the uncertainty of the neural network model. To solve the above problems, an improved feature selection method based on information entropy and feature semantic uncertainty is proposed in this paper. The former is used to obtain fewer and higher quality feature points, while the latter is used to correct the uncertainty of the network in feature selection. At the same time, in the initial stage of feature point selection, this paper first filters and eliminates the absolute dynamic object feature points in the a priori information provided by the feature point semantic label. Secondly, the potential static objects can be detected combined with the principle of epipolar geometric constraints. Finally, the semantic uncertainty of features is corrected according to the semantic context. Experiments on the KITTI odometer data set show that compared with SIVO, the translation error is reduced by 12.63% and the rotation error is reduced by 22.09%, indicating that our method has better tracking performance than the baseline method.

This paper introduces a finite volume element method for solving the time-tempered fractional advection–diffusion equation with weakly singular solution at initial time (t=0). An innovative approach is proposed to construct an intelligent non-uniform temporal mesh, which significantly reduces errors as compared to using a uniform temporal mesh. The error reduction is quantified in terms of percentage improvement of errors. Due to the presence of a large number of integral calculations involving complicated functions, we used parallel computing techniques to accelerate the computation process. The stability of the method is rigorously proven, and numerical examples are provided to demonstrate the effectiveness of the method and validate the theoretical results.