Real-Time Deep-Learning Based Traffic Volume Count for High-Traffic Urban Arterial Roads

Abstract

Traffic volume survey is important for the relevant authorities in estimating road usage and traffic trends for short and long-term traffic facilities planning and design. Commonly, the survey is done manually where the human observers have to be at the actual site throughout the survey period. Not only the method may cause danger to the observers, but it also resources intensive as the traffic volume is increasing, such as in urban arterials. Thus, in this paper, a deep-learning-based traffic volume count system is proposed and extensively tested with 48 high-traffic video clips captured from cameras temporarily installed at four selected urban arterial roads (estimated AADT more than 50,000 and 100,000). For testing, the video clips are split into 5-minutes and 15-minutes duration. Then the accuracy of each clip is evaluated based on the error between system output and the manual ground-truth. The average accuracy for the four camera views is 97.68%, 93.84%, 97.7%, and 94.23% respectively. The system is also able to run in real-time with the average processing time of 37.27ms per frame. Thus, the proposed system is suitable to be used in the traffic volume survey.