A Deep Learning–Based Approach for Moving Vehicle Counting and Short-Term Traffic Prediction From Video Images

Abstract

The intelligent transportation system (ITS) is one of the effective solutions to the problem of urban traffic congestion, and it is also one of the important topics of smart city construction. One particular application is the traffic monitoring and flow prediction. However, there are still challenges regarding both aspects. On the one hand, the current traffic monitoring relies heavily on the single object detection method that cannot achieve accurate statistics of moving target counting and, meanwhile, has limited speed advantage; on the other hand, the existing traffic flow prediction models rarely consider different weather conditions. Therefore, the present article attempts to propose a packaged solution, which combines a new target tracking and moving vehicle counting method and an improved long short-term memory (LSTM) network for traffic flow forecast with weather conditions. More specifically, the DCN V2 convolution kernel and MultiNetV3 framework are used to replace YOLOv4’s conventional convolution kernel and backbone network to realize multi-target tracking and counting, respectively. Subsequently, combined with the temporal characteristics of historical traffic flow, this article introduces weather conditions into the LSTM network and realizes the short-term prediction of traffic flow at the road junction level. This study carries out a series of experiments using the real traffic video data with a 2-month time span at a popular road junction in the downtown of Shenzhen, China. The results suggest that the proposed algorithms outperform the previous methods in terms of the 10% higher accuracy of target detection tracking and about a half reduction of traffic prediction error, when considering weather conditions.