Efficient Edge Computing Device for Traffic Monitoring Using Deep Learning Detectors

Abstract

This article presents a smart camera device for traffic monitoring at intersections. The device is based on the Nvidia Jetson Nano, a small form factor, efficient artificial intelligence (AI) computational device that is capable of deep learning inference. The state-of-the-art deep learning detection models were investigated, and the full YOLOv4 was selected for deployment on the edge device. The deployed model and analytics achieved an average frame rate of 7.8 frames/s (fps). A fisheye lens and camera were selected and integrated with the Jetson processing unit. The original YOLOv4 performed less optimally on fisheye-distorted images. Therefore, we applied transfer learning to the YOLOv4 model using data collected from a local intersection. The final models were evaluated in three different use cases detecting different types of road objects, achieving 100% precision and around 90% accuracy when detecting road vehicles in real time. This article demonstrates the feasibility of running large deep learning models for traffic monitoring services, even on resource-restrained AI edge devices.