KJCKorea
Journal Central

As the edge AI accelerator market rapidly expands, selecting optimal hardware from diverse heterogeneous platforms for real-time object detection has become a critical challenge. However, TOPS (Tera Operations Per Second) figures provided by manufacturers represent only theoretical maximums and fail to reflect real-world performance. This study empirically compares inference performance and cost efficiency for four low-cost heterogeneous edge AI accelerators: Jetson Orin Nano Super, Jetson Orin NX, Hailo-8 M.2, and Rockchip RK3588, using YOLO-series object detection models. Experimental results demonstrate that Hailo-8 achieved 101.2 FPS for YOLOv8s, approximately 4.6 times faster than Jetson platforms. RK3588 achieved 34.6 FPS for YOLOv8n, outperforming Orin NX and Orin Nano Super. In cost efficiency (measured as FPS per USD), RK3588 and Hailo-8 showed 2.2-7.8 times better performance than Jetson platforms. Notably, YOLOv10n with NMS-free architecture exhibited poor performance on RK3588 NPU, highlighting the importance of model-accelerator compatibility. This study provides a cost-efficiency metric framework for edge AI platform selection.

Red tide outbreaks cause severe damage to marine ecosystems, necessitating robust early warning systems. However, traditional models based on single data sources struggle with the nonlinear complexity of environmental factors. This study proposes a multimodal prediction model integrating GOCI satellite imagery and oceanographic numerical data. By combining spatial-temporal satellite features with time-series environmental variables, the model predicts red tide occurrences across 31 Korean coastal regions. The proposed model achieved a Precision of 0.64, Recall of 0.82, and F1-score of 0.72, significantly outperforming existing LSTM-based methods (F1-score = 0.574). These results demonstrate the potential of multimodal integration for real-time red tide forecasting and decision support.

This paper comparatively analyzes depth sampling strategies and their effect on coordinate transformation consistency in a ROS-based RGB-D perception pipeline. The proposed pipeline consists of RGB-based ROI detection, window-based depth sampling for 3D position estimation, ROS TF coordinate transformation, and TCP-based control command delivery. Three sampling strategies (1×1 single-pixel, 3×3 and 5×5 window median) were evaluated across nine experimental conditions, including three camera-to-object distances (30, 40, 50 cm) and three lighting levels (bright, normal, dim), totaling 270 repeated trials. The 3×3 window median improved consistency by 5.7 times over single-pixel sampling, and 5×5 by 8.1 times (Kruskal-Wallis H=71.59, p<0.001). All conditions maintained deviations below 1.3 mm, confirming that depth sampling strategy is a critical design parameter for coordinate transformation reliability.