As the size and complexity of artificial intelligence models increase, the need for distributed computing to efficiently handle large-scale computation and data transfer is becoming increasingly important. However, in a hardware environment that does not support peer-to-peer communication, data transfer between accelerators is limited, which makes it difficult to implement distributed computing. In this paper, we analyse the performance of data transfer between multiple neural processing units (NPUs) based on Peripheral Component Interconnect Express (PCIe) using the Neubla Antara NPU and propose two transfer methods to solve this problem. The first is a Base Address Register (BAR)-based transmission method, in which the central processing unit (CPU) copies and transmits data between NPUs. The second is a CPU-direct memory access (DMA) cooperative transmission method, in which data transmission between the device and the host is handled by DMA, and the host transfers data in stages by the CPU. The results of the experiment showed that the BAR-based transmission recorded a speed of 24.30 MB/s, while the CPU-DMA cooperative-based transmission recorded a speed of 1988.28 MB/s. This study experimentally compares and analyses the performance of the two transmission methods and presents a solution that can support data transmission in an environment where peer-to-peer communication is not possible.

For safe urban driving in autonomous driving platforms, it is critical not only to develop high-performance object detection methods but also to build a training dataset that accurately reflects the diverse environments and object characteristics present in urban settings. To tackle these challenges, in previous study, we created a multi-class 3D LiDAR dataset that incorporated various urban environments and object types. In this paper, we have developed an efficient 3D semi-supervised object detector based on a dual-teacher framework. In this framework, similar classes are grouped into categories, with each category assigned a teacher. By leveraging these teachers, the student model gradually improves, resulting in an efficient object detector. The experiments on the WOD and KITTI validate the effectiveness of our proposed method, and the results demonstrate that our approach consistently outperforms existing state-of-the-art 3D semi-supervised object detection methods.

This study explores the implementation of an efficient golf cart navigation system using deep learning techniques based on camera footage. In constrained environments such as those encountered by golf carts, the accuracy and processing speed of object detection models are crucial. While conventional object detection models offer high accuracy, their computational complexity makes them unsuitable for real-time deployment in embedded systems. In this paper, we introduce Efficient DETR (eDETR), a lightweight object detection model specifically designed to address the challenges of limited computational resources in golf course navigation. eDETR is built upon the Detection Transformer (DETR) architecture and incorporates model compression techniques, including a non-parametric pooling-based encoder and knowledge distillation methods. These enhancements effectively reduce the model's computational complexity while maintaining high accuracy. Experimental results, obtained using a custom dataset collected from an actual golf course environment, demonstrate that eDETR, with the application of knowledge distillation, achieves an Average Precision (AP) of 66.6%, reflecting a 10% reduction compared to DETR. However, the model's computational cost is reduced by approximately 58%, and its inference time on a CPU is increased by nearly quadruple, making it well-suited for deployment in resource-constrained environments. These findings underscore the potential of eDETR as a practical model for real-time object detection in applications analogous to deep learning-based autonomous golf carts.

Early-stage fires appear small in captured images, making accurate detection of small objects crucial for rapid response. Although convolution, pooling operations in convolutional neural networks enhance high-level target recognition, they can also lead to the loss of fine-grained details, thereby degrading the representation and detection performance of small objects. In this paper, we propose a multi-stage detection framework "U-FRCNN (Upscaled-Faster R-CNN)" that combines Faster R-CNN with image upscaling to enhance low-confidence score regions. Initially, objects are detected with Faster R-CNN; then, regions with low detection scores are upscaled and re-evaluated using the same model. Experiments on a dataset from Roboflow and manually labeled images show that our approach improves small object mAP from 0.0947 to 0.1510 (a 59.4% increase) and overall mAP from 0.2183 to 0.2483 (a 13.7% increase), while large object detection remains robust. These results demonstrate the potential of our method for effective early fire detection.

Recent advancements in deep learning technology have gained significant attention in various industrial applications, particularly in the fields of object detection and segmentation. Deep learning-based semantic segmentation techniques offer a powerful tools for precisely identifying and classifying critical elements during driving scenarios. Country club environments present many challenges for autonomous driving systems due to complex terrains, diverse weather conditions, and varying object sizes. As autonomous driving technologies continue to evolve, the ability to accurately perceive and understand these environments becomes essential for safe navigation. In this paper, we propose a LoRA-based cross attention transformer for semantic segmentation of country club driving environments. LoRA enables deep learning models to adapt new environments, and cross attention modules strengthen the flow of information between feature maps. Experiments were conducted on the real world dataset obtained from multiple country club environments. Experimental results demonstrate that our proposed model is superior to existing models for segmentation of country club driving environments. The results of this study are expected to contribute to the advancement of autonomous driving systems and promote further research and development in related fields.

We propose a lightweight algorithm for detecting driver drowsiness by analyzing eye blinking patterns, eye closure duration, and head tilt using the Eye Aspect Ratio (EAR) combined with average central point detection. While EAR effectively detects drowsiness through sustained low values, it can result in false positives during activities such as talking or using a smartphone. To address this, we incorporated gaze stability analysis by comparing the recent average central point of the face with its current position to determine forward attention, enhancing accuracy and reliability in detecting both drowsiness and inattentive behaviors. This approach, combined with Google's MediaPipe Mesh model-which extracts 468 facial landmarks and operates efficiently on mobile and low-spec hardware-ensures stable and highly ac urate detection. Built on the BlazeFace framework with an average precision of 98.61%, this system provides a practical, resource-efficient solution for real-world driver monitoring without the ne d for high-performance computing.