Preprint

Applying 3D human pose and body shape details from a single monocular image presents a significant challenge in computer vision. Traditional methods that rely on RGB images often face constraints due to varying lighting conditions and occlusions. However, advancements in imaging technologies have introduced new techniques, such as single-pixel imaging (SPI), which can overcome these limitations. SPI is particularly effective in capturing 3D human pose in the near-infrared (NIR) spectrum. This wavelength can penetrate clothing and is less affected by lighting variations than visible light, providing a reliable means to accurately capture body shape and pose data, even in challenging environments. In this work, we explore using an SPI camera operating in the NIR range, with Time-of-Flight (TOF) technology at wavelengths of 850-1550 nm, as a solution for detecting humans in night-time environments. Our proposed system employs SPI for depth estimation and feature extraction in humans. These features are then used to generate point clouds, which are integrated into a 3D body model (SMPLX) via 3D body shape regression. This process utilizes deep learning techniques based on self-supervised 3D human mesh methodologies. To evaluate the efficacy of NIR-SPI 3D image reconstruction, we constructed a laboratory scenario simulating night-time conditions. This setup allowed us to test the feasibility of using NIR-SPI as a vision sensor in outdoor environments. By assessing the results obtained from this setup, we aim to demonstrate the potential of NIR-SPI as an effective tool for detecting humans in night-time scenarios and accurately capturing their 3D body pose and shape, with future applications in environmental rescue.

In recent years, Unmanned Aerial Systems (UAS) have become vital assets for Search and Rescue (SAR)  missions, ushering in a technological revolution, due to their versatility, mobility, and accessibility have transformed SAR operations. This comprehensive study explores the current state of UAS deployment in SAR, emphasizing the integration of multi-sensor technologies, categorization of UAVs, payload assessment, and the potential of Artificial Intelligence (AI) in future applications. These applications encompass diverse tasks, from rapid aerial surveys to detecting survivors and hazards in expansive areas. UAS also serve as lifelines for delivering essential supplies to remote regions, aiding communication in areas with poor network coverage, and supplying real-time environmental data for informed decision making. While challenges like regulatory compliance, payload limitations, and battery life

persist, the future of UAS in SAR holds great promise. Advancements in AI and autonomy will empower drones to autonomously tackle intricate missions. UAS systems will amplify coverage and search efficiency, and advanced sensors, including infrared-thermal imaging, LiDAR, RADAR, and cellphone signal detection, will enhance situational awareness. Drone swarms could establish ad-hoc communication networks and conduct extensive searches, while medical payloads offer immediate assistance in remote locations. Machine learning and AI will further refine image recognition for survivor and hazard identification. UAS technology remains a compelling solution to enhance SAR operations, necessitating ongoing research and development efforts to overcome existing challenges and fully unlock their potential in future applications. Adherence to evolving regulations is vital for the safe and effective
integration of these remarkable tools into SAR missions.

The integration of Dynamic Mode Decomposition (DMD) control and Physics-Informed Neural Networks (PINNs) to enhance UAV quadcopter control systems represents a novel approach that combines DMD techniques with PINNs to solve the Riccati equation is crucial for accurate UAV position estimation. This approach formulates the UAV control problem using physics-based constraints, ensuring that the learned models conform to the physical laws governing UAV dynamics. By leveraging a comprehensive dataset that includes UAV flight information such as position, velocity, and control inputs, DMD is utilized to extract fundamental dynamic modes from the data. These modes provide a reduced-order representation that captures the dominant UAV dynamics. This representation is subsequently used within the PINN framework to solve the Riccati equation accurately. Integrating DMD and PINNs creates a powerful control strategy that enhances position estimation accuracy and optimizes control performance. Experimental results validate the effectiveness of the proposed method in real-time UAV control scenarios, demonstrating significant enhancements in estimation accuracy and control stability compared to traditional approaches.

Single-pixel imaging (SPI) has emerged as a promising solution for computational imaging, especially in challenging scenarios where conventional commercial vision systems are inadequate. SPI excels in situations with low light conditions and scarce high-quality cameras in spectral regions. By using structured illumination, SPI can transform one-dimensional signals into high-quality two - and three-dimensional images. However, the reconstruction quality of SPI images heavily relies on the number of temporal domain measurements. Deep learning has the potential to improve SPI image reconstruction, but it faces difficulties due to limited training data, particularly for specific wavelength bands. In this research, we present an innovative approach that combines the Deep  Image Prior method with Generative Adversarial Networks (GANs) to enhance the quality of SPI images captured in the near-infrared (NIR) range (850 to 1550 nm). This method eliminates the need for direct SPI image datasets by utilizing an unsupervised image super-resolution approach based on Deep Image Prior. We evaluate various neural network configurations, employing UNet and GAN architectures, to determine their effectiveness in SPI computer vision systems.

​OFDM-based radar systems have demonstrated capabilities comparable to traditional Linear frequency-modulated (LFM) radar systems in range and velocity estimation, and image generation. These systems find applications in diverse fields ranging from traffic monitoring to unmanned vehicles and robotic vision. The orthogonality of OFDM signals allows simultaneous transmission from multiple antennas using spectral division, enhancing system efficiency. The adoption of OFDM in radar systems brings significant enhancements over LFM-modulated systems, particularly in spatial resolution. This is achieved through space diversity using MIMO array antennas, controlled bandwidth, and reduced range ambiguity by incorporating a cyclic prefix (CP). The integration of OFDM in microwave imaging systems significantly improves resolution and aperture synthesis. This paper introduces the development of a cost-effective orthogonal frequency division multiplexing (OFDM) radar system, utilizing multi-carrier in-band S technology in a 1 x 2 MIMO configuration. It focuses on generating 2D images using the Range Doppler Algorithm (RDA), with enhanced spatial resolution through carrier control and phase adjustment in the OFDM system. The proposed radar system is tailored for unmanned ground vehicle (UGV) applications. It employs a virtual antenna technique to extend the antenna aperture, thereby covering a more effective measurement area.