Real-Time Traffic Status Detection from on-Line Images Using Generic Object Detection System with Deep Learning

Abstract

In order to leverage the traffic burden and reduce traffic time, the Macao Government provides instant on-line traffic information for road users. In this paper, we propose a unified low-cost real-time traffic status detection system using free on-line images. Since the quality for vehicle detection from these on-line images can vary very differently, depending on the locations of cameras and weather conditions, we propose to use a generic object detection system with deep learning, e.g. YOLOv3, to detect real-time traffic status from these images. In order to determine the traffic status using the predicted bounding boxes and the number of detected vehicles output from the YOLOv3 detector, we propose to use the intersection over union (IOU) of the union of the bounding boxes for multiple objects predicted on images taken at different times, termed mIOU, and the corresponding estimated number of vehicles to estimate the multi-level traffic status. From the experimental results, it was found that the precision and recall rate can achieve 86% and 87%, respectively, depending on the locations of the cameras. It was also shown that the proposed system using YOLOv3 and mIOU can be used for traffic status estimation for low-resolution images, without requiring any pre-training of networks for different settings. The proposed traffic detection technique is applicable for a wide range of application scenarios.