Improved Real-Time Traffic Congestion Detection with Automatic Image Cropping using Online Camera Images

Traffic congestion causes severe traffic delays, increases fuel wastage and monetary losses. Real-time information on traffic congestion can help road users avoid congested road segments and thus reduce commute time. In this paper, we propose an improved version of the real-time traffic congestion detection system using online camera images. To improve the detection accuracy, we propose an automatic image cropping method to identify the region of interest, based on the detected locations using the bounding boxes produced by YOLOv4. We improve the use of instantaneous intersection over union for multiple bounding boxes (mIOU) values to simplify the process of traffic status estimation, by taking the weighted average of the current and the previous instantaneous mIOU values. The experimental results show that it is feasible to define a congestion detection threshold for weighted mIOU when the time interval between the two images is greater than 2 seconds. With a threshold of 0.3, the average Precision and the average Recall Rate can achieve 89.84% and 92.09% in the manually labeled test data sets. The detected numerical traffic status results can be used for short-term traffic prediction.