KJCKorea
Journal Central

In this paper, we propose a multiprocessing-based parallel pipeline architecture for robust real-time implementation of multi-target designation and tracking algorithms in embedded environments. While the demand for on-site complex AI vision is surging in advanced industries, real-time processing on resource-constrained embedded systems remains a significant technical challenge. The proposed architecture addresses this by separating pre-processing, inference, and post-processing into independent processes, enabling frame-level parallelism. Furthermore, it efficiently distributes computations across CPU, GPU, and NPU resources to minimize idle time and improve overall throughput. On the Jetson AGX Orin platform, the architecture achieves an 88% improvement in FPS compared to a single-process baseline with stable enhancements in tracking accuracy. In contrast, on a server using the TensorRT runtime, performance gains were limited due to conflicts with its internal optimizations. These findings demonstrate that our architecture enables the on-device implementation of complex vision AI models without high-performance servers, making it a key technology for applications such as precision-guided weapons and autonomous drones.

Currently, there is a need to develop a monitoring transceiver terminal that provides an environment for applying and testing algorithms capable of predicting and inferring the expected movement paths and operation times of various moving objects. In this study, a GNSS-based RTLS/IoT location tracking and monitoring terminal was developed by applying a GNSS-INS integrated module to the hardware, combining GNSS data with inertial sensing and velocity/heading sensor information to deliver precise positioning for vehicles and objects. Utilizing this system enables the development and optimization of algorithms for detecting optimal movement paths, ensuring stable management of moving objects, and handling large volumes of logistics within limited timeframes while enhancing algorithm scalability. Furthermore, the developed terminal can be widely utilized in intelligent mobility applications such as autonomous vehicles, unmanned shuttle buses, taxis (e.g., Uber), and freight transport, as well as in unmanned patrol and exploration vehicles for hazardous areas to ensure human safety. It is also applicable in everyday domains such as unmanned indoor and outdoor cleaning systems.

Multi-document Question Answering (Multi-document QA) is an essential component of question answering systems in real-world environments. It requires extracting and integrating the necessary information from multiple documents to generate accurate responses. The DS-RAG framework is optimized for real-world multi-document QA tasks. It improves the system’s overall accuracy and efficiency by using a document selection module that selects only the core documents retrieved after question decomposition, which are necessary for response generation. However, this process requires additional computational resources and additional training, making it difficult to apply in resource-limited environments. To address this limitation, this study proposes a Lightweight DS-RAG framework that achieves selection effects through importance-weighted query generation, thereby eliminating the need for a separate selection module. The proposed approach identifies key information within a question and generates importance-weighted queries. This enables retrieval focused on the most relevant information while maintaining both efficiency and accuracy. This approach maintains high retrieval accuracy while significantly reducing computational overhead, and can be applied in resource-constrained environments without domain-specific training. Experimental results demonstrate that the Lightweight DS-RAG framework achieves 95% of the retrieval performance of the original DS-RAG, as measured by the F1-score. Additionally, it exhibited an average performance improvement of 67.5% over RAG without a selection module. These results demonstrate that a high level of accuracy can be maintained without a separate selection module or additional training, indicating that the Lightweight DS-RAG framework can serve as a practical alternative in resource-constrained environments.