Digital twin systems have been increasingly adopted in the manufacturing industry with recent advancements in IoT, deep learning, and physical AI technologies. These systems are improving productivity for factory automation and process optimization, and are also applicable to safety management within manufacturing environments. In particular, unexpected collisions between workers and robots can be occurred in human-robot collaborative work cell. To prevent the collisions, it is essential to recognize workers in 3D within the workspace and simulate them in a digital twin environment to predict the potential collisions. In this study, we developed a multi-view custom dataset captured with a thermal infrared camera to analyze and even in environments where visibility is impaired by factors such as dust or smoke. We applied video object segmentation and novel view synthesis models to simulate the worker as a 3D point cloud. The performance of models was quantitatively evaluated to identify those best suited for our proposed method. Our results demonstrate that the proposed approach can successfully reconstruct workers as 3D point clouds in robot work cell. This contributes to enhancing the safety and efficiency of smart manufacturing and surveillance applications.

3D Gaussian Splatting has emerged as a prominent technique for realistic 3D avatar creation, owing to its high-quality real-time rendering capabilities. However, reconstructing head avatars from monocular videos remains challenging due to geometric artifacts caused by depth ambiguity. Furthermore, existing texture-based Adaptive Density Control (ADC) methods often generate excessive Gaussians even in geometrically simple regions, leading to inefficiencies in memory and rendering. To address these limitations, we propose an efficient 3D head avatar reconstruction method utilizing Curvature-based Adaptive Density Control. Our approach calculates a curvature map that quantifies geometric complexity from normal vectors estimated from monocular inputs and integrates this metric into the density control process to selectively distribute Gaussians in geometrically significant areas. This strategy effectively suppresses unnecessary Gaussian generation in flat regions while concentrating resources on detailed features such as facial landmarks and wrinkles, thereby maximizing representation efficiency. Experimental results on the INSTA dataset demonstrate that the proposed method achieves high-fidelity rendering quality visually comparable to state-of-the-art baselines while using fewer Gaussians. Notably, our method improves the quality-per-Gaussian efficiency score by up to 13.4% in quantitative evaluations.

This paper proposes a robust hybrid Stanley-ADRC control strategy for the precise waypoint tracking of tracked solar pile drivers in unstructured environments. By utilizing an extended state observer (ESO), the proposed method estimates and compensates for model uncertainties and external forces as "total disturbance" in real-time, overcoming the limitations of conventional linear controllers. Comparative simulations demonstrate that while the baseline Stanley-PID system failed to track the path due to actuator saturation caused by accumulated errors, the proposed Stanley-ADRC framework effectively cancelled out disturbances and achieved stable tracking. These results verify that the proposed framework secures the robustness and accuracy essential for automated operations in disturbance-prone conditions.

This paper presents the development of an FPGA-based automated testing simulator for Transmit/Receive (TR) modules deployed in next-generation Synthetic Aperture Radar (SAR) satellites. TR modules, which control antenna phase and gain for microwave signal transmission and reception, are core components that are required to operate reliably in harsh space conditions. Conventional testing methods rely on chamber-based testing, which is time-consuming, costly, and difficult to scale. Furthermore, environmental stress can induce anomalies, requiring hardware redesign and repeated testing and validation, significantly increasing development time and cost. To address these challenges, we propose an FPGA-based simulation platform that enables early design validation and automated performance testing under simulated environmental conditions. The system implements real-time anomaly detection and incorporates automated validation routines compliant with MIL-STD-883, a widely adopted standard for qualifying spaceborne components. Experimental verification confirms that the simulator significantly improves test efficiency and reliability while reducing development time and cost compared to traditional methods. The results demonstrate the practical effectiveness of the simulator as a tool for accelerating the design, verification, and qualification of space-grade TR modules.

Typical embedded system applications are implemented in the form of calling task functions according to a fixed schedule. In such embedded software development environments, debugging is often cumbersome. Whenever a task schedule changes or a task requires modification or contains errors, developers must repeatedly compile the code on the host system, download the firmware, and test the tasks. Enabling extensive task testing after a single firmware download can therefore greatly improve development productivity. This paper proposes an interactive development environment framework for embedded systems based on a non-preemptive FCFS scheduling scheme, which allows tasks to be executed and tested through console commands-similar to general-purpose operating systems such as Linux. Experiments on an ATmega2560 MCU confirm that tasks can run both on a fixed schedule and in response to console requests.

This work proposes enhancing image classification performance through heterogeneous trajectory learning rates and weight fusion. Two models are independently trained using exploratory trajectory learning and local trajectory learning, then their weights are averaged to initialize re-adaptive optimization. The exploratory trajectory explores broad parameter space to avoid local optima, while the local trajectory ensures stable convergence. This weight fusion strategy leverages distinct optimization paths, guiding the model toward flatter loss regions that improve in-distribution test accuracy. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet with ResNet, VGG, AlexNet, and ViT architectures show consistent improvements over baselines, with ResNet34 achieving a 2.92% gain on CIFAR-100. These findings demonstrate that this framework effectively enhances model performance without additional inference costs.

The use of edge device-based object classification technologies is expanding in areas such as smart farming, automated harvesting robots, and distribution automation. As a result, the demand for lightweight deep learning models that can support efficient real-time processing is increasing. This study evaluates the performance of three representative lightweight models-MobileNetV4 (Mobile Network Version 4), SHViT (Single-Head Vision Transformer), and ViM (Vision Mamba)-for classification of agricultural products. Experiments were conducted in both a high-performance computing environment with an NVIDIA RTX 4070 GPU and an edge device environment using the Jetson AGX Orin platform. The models were compared in terms of Top-1 accuracy, MACs (Multiply-Accumulate Operations), FPS (Frames Per Second), and NetScore. The results show that MobileNetV4 achieved the best overall balance of accuracy, efficiency, and processing speed. These findings confirm the feasibility of applying lightweight deep learning models to real-time object classification in resource-constrained edge environments and provide useful insights for future applications in agriculture and industrial automation.