Preprint

Applying 3D human pose and body shape estimation from a single monocular image is a significant challenge in computer vision. Traditional methods relying on RGB images often struggle with issues caused by varying lighting conditions and occlusions. However, advancements in imaging technologies, such as single-pixel imaging (SPI), have introduced new techniques that overcome these limitations. SPI is particularly effective for capturing 3D human poses in the Near-Infrared (NIR) spectrum. NIR light can penetrate clothing and is less sensitive to lighting variations than visible light, making it a reliable method for accurately capturing body shape and pose data, even in challenging environments. This work explores using an SPI camera operating in the NIR range, combined with Time-of-Flight (TOF) technology, at 850-1550 nm wavelengths 850-1550 nm to detect humans in low-light or night-time environments. Our proposed system leverages SPI for depth estimation and human feature extraction. These features
generate point clouds representing human poses, which are then used for 3D body shape regression through a self-supervised network based on PointNet++. This network estimates parameters such as global rotation R, translation t, body shape β, and pose θ using the SMPLX model to generate a 3D human mesh reconstruction. We constructed a laboratory scenario simulating night-time conditions to evaluate the efficacy of NIR-SPI 3D image reconstruction. This setup allowed us to test the feasibility of using NIR-SPI as a vision sensor in outdoor environments. By analyzing the results, we aim to demonstrate the potential of NIR-SPI as an effective tool for human detection in night-time scenarios, enabling precise 3D body pose and shape capture with future applications in environmental rescue and other fields.

In emergencies, particularly in disaster zones, detecting RF signals is critical in locating victims. These signals, often from various communication sources, are identified by systems scanning across multiple frequencies to aid rescue teams. In this work, we present a cost-effective detection system utilizing RTL-SDR hardware integrated with FPGA technology, capable of identifying 10 distinct types of radio signals, including cellular, radio broadcast, and satellite communication modulations, commonly encountered in such scenarios. We implemented three deep neural network architectures to enhance signal detection, integrated with FPGA-based DPUs alongside RTL-SDR hardware. The best-performing network achieved an impressive accuracy of up to 98%, positioning this system as a promising candidate for future UAV-based rescue operations.

Unmanned Aerial Systems (UAS), commonly known as drones, have become essential assets in Search and Rescue (SAR) operations due to their versatility, rapid deployment, and high mobility. This study reviews drones' current and emerging uses in SAR, with a focus on advancements in sensor integration, payload capacity, and multi-UAV coordination. It further explores how Artificial Intelligence (AI) can enhance operational efficiency. A comprehensive review methodology is employed to analyze recent progress in drone technology, AI, and digital twin simulations aimed at optimizing SAR missions. The review highlights the current capabilities, strengths, and limitations of existing systems while identifying potential innovations to address persistent challenges. Drones are now effectively deployed to survey vast areas, locate survivors, and assess hazards. Coordinated multi-drone systems have the potential to expand coverage, enhance efficiency, deliver essential supplies, and establish temporary communication networks in inaccessible regions. Future advancements in AI and autonomy will enable drones to perform complex tasks with minimal human intervention. Enhanced sensor technologies will improve detection capabilities, including infrared imaging, radar, and biometric monitoring. However, challenges such as regulatory restrictions, limited battery life, and payload constraints persist. Addressing these challenges will require ongoing research and technological breakthroughs. This study underscores the transformative potential of evolving drone technologies in SAR operations, paving the way for faster, more efficient responses, ultimately saving lives through improved real-time decision-making and operational capabilities.

In recent years, Physics-Informed Neural Networks (PINNs) have emerged as a promising approach for addressing complex inverse problems in image processing, particularly in the realm of image denoising. This paper presents an innovative framework that leverages various neural network architectures, including ResUNet, UNet, U2Net, and Res2UNet, to effectively implement denoising strategies based on nonlinear partial differential equations (PDEs). The proposed methodologies utilize PDEs such as the heat equation, diffusion processes, multiphase mixture and phase change (MPMC) models, and the technique developed by Zhichang Guo, integrating physical laws into the learning process to enhance the robustness and accuracy of denoising. We demonstrate how these models can be effectively trained to minimize noise while preserving essential image features, utilizing a combination of data-driven approaches and physical constraints. Our experiments reveal that incorporating PDEs significantly improves denoising performance compared to traditional methods. The models are evaluated on various datasets, and performance metrics such as the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) indicate substantial improvements in image quality. The results underscore the efficacy of using PINNs with nonlinear PDEs for advanced image-denoising tasks, paving the way for future research in the intersection of deep learning and physics-based modeling in image processing applications.