Optimized YOLOv8 (YOLO-Prime), M-CNN, and Dynamic Linear Regression for Real-Time Aerial Vehicle Detection in Embedded Systems

Abstract

This paper introduces a modified architecture of YOLOv8, called YOLO-Prime, to detect aerial vehicles. By employing YOLO-Prime to embedded systems such as Nvidia Jetson Xavier Nx and Rockchip 3588, we reduced the inference time by 130% compared to the YOLOv8m model and by 158% compared to the YOLOv8l model. In addition to YOLO-Prime, we present a Morphological CNN (M-CNN) to classify regions detected by YOLO-Prime. With the integration of M-CNN, we achieved an increase in detection accuracy by 20% in the detection of aerial vehicles. Furthermore, we introduce a dynamic linear regression model to predict the future coordinates of aerial vehicles. As a result of the integration of dynamic linear regression models to embedded systems, GPU utilization of deep learning models decreased by approximately 66%. In summary, the combination of YOLO-Prime, M-CNN, and the dynamic linear regression model contributes to detecting aerial vehicles in embedded systems where energy consumption, inference time, and accuracy are crucial for real-time applications.