In this study, the method used involves segmenting images into fixed-sized square regions called superpixels and then analyzing features of each superpixels to determine whether it belongs to the foreground or background. The study also discusses the process of selecting relevant features and testing different combinations. These features include characteristics of the superpixel itself—its local relationships with neighboring superpixels, and global features of the entire image. Importantly, it emphasizes making the decision-making process understandable—achieved by using a decision tree model for analyzing the collected features. This study focuses on teaching a self-driving system how to accurately separate foreground (movable areas like floors) from background (non-movable areas) in images captured inside buildings. This approach aims to strike a balance between entirely automated segmentation models like Convolutional Neural Networks (CNNs)—which lack explainability, and purely knowledge-based segmentation models, where all information is predefined. This information is stored in a database for analysis. Specifically, it addresses the scenario of an autonomous vehicle navigating public spaces without endangering people.

This paper presents an in-depth exploration of scalable autonomous driving solutions leveraging advanced computer vision techniques. As the demand for autonomous vehicles (AVs) grows, ensuring scalability across diverse environments and vehicle platforms becomes a critical challenge. The proposed solutions integrate state-of-the-art computer vision algorithms with machine learning models to enhance the perception, decision-making, and control mechanisms essential for autonomous driving. We explore the use of convolutional neural networks (CNNs), real-time object detection, semantic segmentation, and 3D scene reconstruction to create a robust, adaptable framework capable of handling varying traffic conditions, road types, and weather scenarios. Additionally, we address the scalability of these solutions in terms of computational efficiency and the ability to generalize across different autonomous driving systems. Through extensive simulation and real-world testing, our results demonstrate significant improvements in accuracy, responsiveness, and overall safety, paving the way for the widespread adoption of autonomous driving technologies.

This study presents the world's fastest three-dimensional double-directional channel sounding system operating at 105 GHz with a null-to-null bandwidth of 4 GHz. The developed channel sounding system allows a full scan of the 105 GHz multipath channel in both azimuth and elevation angles of the transmitter (TX) and receiver (RX) directional antennas. The developed channel sounding system is calibrated with an anechoic measurement in that the path loss of the observed line-of-sight ray coincides with the free-space path loss. Moreover, the anechoic measurement verifies the performance of the channel sounding system, where the delay resolution is lower than 0.5 ns, and the angular resolution is approximately 10° in the azimuth plane for both TX and RX.

By capturing gaze points and eye movement patterns, eye tracking technology enables researchers to gain insights into visual attention, cognitive processes, and human behavior. Raw eye tracking data is inherently noisy, containing coordinates with occasional invalid entries. Researchers commonly face the challenge of cleaning and refining this data before proceeding to calculate essential eye tracking metrics. Critical preprocessing steps involve filtering out noise, detecting events such as fixations and saccades, and deriving metrics such as fixation duration, saccade amplitude, and more. This paper introduces a tool that provides users with the ability to preprocess data from both Gazepoint and FOVIO eye trackers. The methodology adopted for this line of work is based on Nyström and Holmqvist's (2010) data and velocity-driven algorithm. Preliminary results show the methodology developed as part of this line of work is promising in terms of improved accuracy and reliability in detecting fixations and saccades. Furthermore, this code and its versatile Graphical User Interface (GUI) has been successfully employed in multiple studies. Finally, we present supplementary code for calculating various eye tracking metrics and percent gaze overlap, extending the utility of our tool for in-depth analysis of visual attention and collaborative gaze behavior.

In CT/PET imaging applications, reconstructing images from low-dose/low-count acquisitions often leads to lower image quality, necessitating specialized denoising methods and reconstruction algorithms to enhance diagnostic accuracy. While many recent denoising techniques employ Convolutional Neural Networks (CNNs), these architectures may struggle with capturing long-range, non-local interactions, potentially resulting in inaccuracies in global structure representation. Recognizing the advantages of transformer architectures over CNNs on that front, our study introduces a novel sinogram denoising algorithm tailored at improving low-dose/low-count sinogram quality. We propose a transformer-based sinogram denoiser module specifically designed to match the structure of sinogram data, enhancing sinogram feature extraction and denoising performance. Furthermore, by incorporating image domain denoising, we propose cross-domain image reconstruction, allowing for further image quality refinement by addressing image-specific noise characteristics. Our cross-domain image reconstruction network, which incorporates the proposed sinogram denoiser module, has been trained with both synthetic and clinical data. Performance evaluations reveal that our Sinogram Sinusoidal-Structure Transformer Denoiser achieves outstanding results in sinogram denoising, while our Cross-Domain Image Reconstruction Network demonstrates excellent image reconstruction capabilities, as validated by both subjective and objective metrics.

This article presents a novel four-port probe calibration method for on-wafer S-parameter measurement and, for the first time, realizes measurement verification up to 110 GHz. The method regards the 64 error terms in the four-port probe calibration as an error matrix, which can be solved by a homogeneous equation system using a generalized scatter matrix theory. The calibration algorithm can consider the crosstalk between the probes and improve the calibration accuracy at high frequencies. This method only requires six calibration standards to complete the four-port probe calibration. Only one coupled differential line with a symmetric structure is needed as the Thru calibration standard, which replaces the traditional U-shape and asymmetric Thru and effectively reduces the number of calibration standards used in the four-port probe calibration. In addition, the method combined with an optimization method accurately and efficiently calculates the parasitic parameters of the calibration standards at high frequencies. The proposed method is validated by four different devices under test (DUTs) up to 110 GHz. Their S-parameters obtained by the proposed method, including all the pure-mode and mode-conversion terms, have high accuracy in magnitude and phase, proving the correctness and convenience of this method.

A document by Habib Ullah Manzoor. Click on the document to view its contents.

The classic implementation of dispatchable virtual oscillator control (dVOC) in the stationary αβ frame is counterintuitive and complex for small-signal modeling of the controller. Additionally, it makes integrating advanced control loops into the dVOC controller challenging and cannot achieve zero reactive power tracking error. To address these issues, for the first time, a new implementation of the dVOC in the dq frame is proposed by transforming the dynamics of the instantaneous angular frequency and voltage magnitude of the dVOC-controlled voltage vector into the local synchronous rotating frame of the converter. This new implementation resolves the inconvenience associated with small-signal modeling, enhances the flexibility and compatibility of the controller, and eliminates reactive power tracking errors, without compromising the operational characteristics of the dVOC. These improvements are achieved through the control of DC variables and the direct regulation of the converter phase angle and voltage magnitude. The effectiveness of the new dVOC controller is validated through EMT simulations.

There is growing tendency toward the application of high frequency electromagnetic waves in in areas such as wireless telecommunications and military technology. High frequency electromagnetic waves like microwave, millimeter wave and terahertz penetrate to the surface layer of a medium depending on the frequency and the conductivity. Moreover, the most important types of interaction between high frequency electromagnetic waves and the surrounding medium - when the medium is represented by a biological tissue model - are predominantly in the form of tissue heating. The objective of this study is to analyze the influence of exposure to electromagnetic fields in the frequency range of 1 MHz – 100 GHz on a biological model coated with soil samples. The biological model consists of three biological tissue: skin, fat and muscle. Further, four soil samples are used as a coating material based on their textual classifications which are silty clay, silty loam, sandy loam and loam. Results show that the different soil samples can reduce the amount of absorbed radiation, and heating in the skin tissue thus proved to be a good shielding material against high frequency electromagnetic waves, particularly for frequencies greater than 10 GHz.

The rapid advancement of Artificial Intelligence (AI) models such as Generative Adversarial Networks (GANs) has been a great success in the field of image synthesis and creation. Artificially generated GAN-based images are widely spread over the Internet along with the development in generation of natural and photorealistic images. While this could lead to better digital media and content, it also poses a risk to security, legitimacy, and authenticity. The advancement of AI-generated images, particularly those that are produced by Generative Adversarial Networks (GANs), has created a rising concern about the potential misuse of these images in spreading misinformation and creating deepfakes. Detecting such fake or AI-generated images has become an important challenge in maintaining the integrity of digital media. In this research, we have explored the application of the Vision Transformer (ViTs) model for detecting AI-generated images, leveraging the Kaggle dataset-a balanced collection of real and AI-generated images. The Vision Transformer is recognized for its innovative method of treating images as sequences of patches and excels at identifying long-range dependencies and complex patterns within images. That makes it exceptionally well-suited for this task of detecting fake images. We have fine-tuned the ViT model on the dataset, performing data augmentation techniques on it and leveraging pretrained weights to boost the model's performance. The findings thus obtained demonstrate that the ViT model attains a high level of accuracy in differentiating between real and AI-generated images, outperforming traditional CNN-based approaches. Beyond performance evaluation, we also conducted an ablation study to examine the impact of various components of the ViT model, including the number of attention heads, patch size, the impact of data augmentation, and the depth of layers. The results obtained in this study indicate that the ViT model not only excels in accuracy but also provides a robust framework for detecting AI-generated images across diverse scenarios. Our study shows the strength of transformer based models in addressing the increasing challenge of AI-generated image detection, laying a foundation for future research in this critical area. This experiment highlights that when the ViT model is fine tuned with optimal data augmentation techniques, it gains state of the art performance in AI-generated image detection, emphasizing its potential for real-world applications.

This paper presents a method for detecting active gas stoves with coil burners using a Raspberry Pi and Pi Camera, focusing exclusively on the signal and image processing components. The system captures images at 30-second intervals to monitor stove activity. Our approach involves processing these images to identify the thermal characteristics of active coils. Although the broader goal was to develop a complete alert system for kitchen safety, this work details only the image processing techniques and their effectiveness in detecting active stoves. The results demonstrate that the proposed method successfully identifies active coils, providing a foundation for future integration of an alert mechanism.

As wireless communication networks evolve to support more devices and higher data rates, ensuring secure communication has become increasingly critical. Physical Layer Security (PLS) offers a promising approach to safeguarding data transmission by exploiting the physical properties of the communication channel. However, implementing PLS techniques often requires significant energy resources, which can be a limiting factor, particularly in resource-constrained environments like IoT networks. Energy harvesting (EH) technology presents a viable solution by enabling devices to harness ambient energy, thereby supporting the continuous operation of PLS mechanisms. This paper explores the role of energy harvesting in enhancing physical layer security, discussing various EH techniques, their integration with PLS strategies, and the challenges and future directions in this emerging area of research.

This paper presents a novel guidance and control strategy for the efficient capture and stabilization of fast-spinning target satellites, such as spin-stabilized models, utilizing a spinning-base servicing satellite equipped with a robotic manipulator, joint locks, and reaction wheels (RWs). The proposed method leverages the RWs of the servicing satellite to replicate the target satellite's spin dynamics, while locking the manipulator joints to achieve precise spin-matching. This technique effectively neutralizes the target's relative motion, simplifying the capture trajectory planning and eliminating the need for post-capture adjustments. Following spin-matching, the manipulator's joints are unlocked, enabling a coordinated control system to execute the capture maneuver while maintaining zero relative rotation between the servicing and target satellites. Upon successful capture, the system transitions into the spin stabilization phase. Here, the joints are re-engaged to create a unified tumbling rigid body composed of the interconnected satellites. An optimized control system then applies negative torques through the RWs to rapidly dampen the tumbling motion, constrained by the actuation limits of the RWs and the manipulator's end-effector torque capabilities.

Most works on reflecting-only intelligent reflecting surface (IRS) are based on the ideal phase shift model (IPSM), which assumes that the reflection amplitude is independent of the phase shift at each IRS element. However, due to hardware limitations of practical IRS circuits, practical IRS generally follows the practical phase shift model (PPSM), where the reflection amplitude is related to the phase shift. This paper analyzes the average block error rate (BLER) of PPSM-based IRSaided short-packet communication (SPC) with statistical channel state information over Rician fading. Because the optimal continuous phase shifts at the PPSM-based IRS are difficult to be explicitly solved, we first propose a deep deterministic policy gradient (DDPG) based algorithm to optimize the phase shift design. Then, we derive the average BLER of the considered IRS-aided SPC with the optimized reflection matrix. Finally, numerical and simulation results demonstrate that the IRS-aided SPC system with the proposed DDPG-based phase shifts outperforms the system with the IPSM-based optimal phase shifts. They also show that, different from the IPSM-based IRS, the optimal phase shifts at the PPSM-based IRS are related to the Rician factors.

This paper presents an Adaptive Spatial-Temporal Convolutional Neural Network (ASTCN) model for traffic state prediction, which learns the complex spatial-temporal information of traffic data through trainable temporal embedding and graph adjacency matrix. The gate control mechanism is used to control the proportion of different temporal embedding to improve prediction accuracy. In the visualization of the temporal embedding, it is found that the influence of the weekly temporal embedding on prediction accuracy is the greatest, followed by the daily temporal embedding, and the monthly temporal embedding is the smallest. Meanwhile, a new method for constructing graph adjacency matrices is proposed, and the experimental results show that the two types of graph adjacency matrices have different effects on improving the prediction accuracy of traffic speed and traffic flow. Therefore, this paper fuses the two graph adjacency matrices to make the ASTCN model achieve better prediction accuracy on both traffic speed and traffic flow datasets than when the two are present separately. The final prediction results of the ASTCN model in the comparative experiment show excellent performance.

This paper proposes an AC Fault Ride-Through (AC-FRT) control for VSC-Based HVDC systems in accordance with modern grid codes. The presented strategy also provides a backup energy controller for AC faults for cross-control MMCs where the converter total energy is normally regulated through the AC side current. The proposed AC-FRT control is capable of positive and negative sequence voltage support and properly limits current set-points to the specified safe limits accordingly with a tunable response time. Furthermore, the backup energy control allows for a distinct response for AC-FRT operation, independent of the energy controller employed during normal operation. The effectiveness of the proposed methods are validated using time-domain simulations in a Multi Terminal Direct Current (MTDC) grid. Furthermore an analysis is carried out to determine the effects of energy control in Modular Multilevel Converters (MMC) and the AC-FRT controls for different fault scenarios on the AC and MTDC grids interlinked by the converter. The aforementioned analyses will be used to attain a heuristic tuning method for the converter's AC-FRT operation. The AC-FRT control's time constant is the tuning parameter in 2 level VSCs where in the case of an MMC, the backup energy controller can also be designed in tandem with the AC-FRT time constant in order to attain a reasonable balance between AC fault current response time and MTDC grid dynamic variations.

In our work, we demonstrate a novel Neural Network (NN) based methodology for generating S-Parameters based on using the meshed and discretized geometry as input to the NN. Our approach would enable a designer to rapidly optimize the performance of vias across an entire substrate. This performance is possible while also achieving a computational time improvement of approximately double that in other NN-based works.

Hardware design for radio frequency devices often demands the use of full-wave electromagnetic (EM) field-solvers to extract scattering parameters (S-parameters) for systems that contain discontinuities throughout their printed circuit boards (PCBs) and substrate packages. Using EM field-solvers not only requires extensive EM knowledge during the setup phase, but also demands considerable computational resources. Previously, machine learning (ML) and neural network (NN) models have been used to solve EM fields; however, they tend to use design parameters as input rather than the structure's geometry. Parametric analysis is beneficial when the solution space is well bounded, but it is more desirable to develop the capability of feeding generic and arbitrary geometry into a machine learning model for rapid S-parameter extraction. This work presents an algorithm for converting a subsection of hardware geometry designed with an electronic design automation (EDA) tool into a dataset suitable for training and validating machine learning models including fully connected NNs and convolutional NNs (CNNs).