Detection of unwanted traffic congestion based on existing surveillance system using in freeway via a CNN-architecture trafficnet

Abstract

Detection of traffic congestion is important for route guidance using in intelligent transport system (ITS) to prevent jam escalation. Although the surveillance system has been used in freeway for years, it is hard to automatically identify and report traffic congestion in complicated transportation scene according to various illumination, weather and other disturbances. The detection process based on human eye is time-consuming and tedious as the machine detection accuracy is not high enough to meet the requirements of practical applications. In this paper, a new classifier is proposed using convolutional neural networks (CNN) to generate four TrafficNet based on two championships of ILSVRC including AlexNet and VGGNet. Instead of using fully-connected layers in AlexNet and VGGNet, a support vector machine (SVM) are used after CNN architecture. Congestion and non-congestion images are trained and tested through this new structure. Image database with more than 30000 images are extracted from existing traffic surveillance video and corresponding labels are added manually. With database, those TrafficNet are trained and tested, detection accuracy and training time of those TrafficNet are compared. The experimental results show that the accuracy of proposed method can reach up to 90%, which is much higher than traditional method based on feature extraction without deep learning.