Real Time Object Detection of Aerial Images Using Deep Learning on Jetson Nano

Abstract

Rapid advancements in technologies such as autonomous driving, humanoid robotics, and vision augmentation have led to a growing need for accurate and quick object identification algorithms that operate in real-time on edge devices. For the purpose of detecting and classifying the different kinds of vehicles (such as trucks and cars) crossing the main road, the system was constructed utilizing the YOLOv8 and Faster RCNN object detection model. This made it possible to develop a small, effective system that can process aerial imagery in real time and is ready for deployment in a variety of situations requiring minimal resources. The main goal is to create a reliable and effective vehicle detection system that works with the Nvidia Jetson Nano card. Within NVIDIA's Jetson range of specialized SOCs for AI systems, the Nano is the smallest model available. It has a computing capacity of 472 GFLOPs with a power consumption of 5 Watts. It offers a CPU and GPU integrated into the SOC. These characteristics are within the range of what one may logically anticipate from a widely accessible edge device. The task was accomplished by making use of a sizable aerial dataset that was gathered with UAVs while paying attention to various circumstances and scenarios. As a result, an effective detection model was trained having low latency and high accuracy for vehicle detection on images captured by UAVs.