Multilevel Traffic State Detection in Traffic Surveillance System Using a Deep Residual Squeeze-and-Excitation Network and an Improved Triplet Loss

Although a substantial number of traffic videos have been accumulated via daily monitoring, deep learning is seldom utilized to process these data for multilevel traffic state detection. The application of deep learning is limited for two reasons: (1) the multilevel traffic state based on traffic images has not been defined. (2) The high noise information in traffic images and extremely similar features of adjacent traffic states hinder accurate detection. Based on this situation, A new definition of the image-based multilevel traffic state is proposed using the ratio of the vehicle areas to the road areas in a traffic image, and a standard image dataset, including various illuminations and vast scenes, are established. A deep residual network named TrafficNet, which is embedded with Squeeze-and-Excitation blocks and is learned by the improved triplet loss, is proposed for multilevel traffic state detection. The Squeeze-and-Excitation block effectively reduces the model's attention to noise information and focuses on road areas that are associated with traffic features in an image. The improved triplet loss maps the learned features to a metric space where the distance between features of inter-class is larger than that within the same class, which improves the discrimination of features between adjacent traffic states. Relevant experiments prove that the performance of TrafficNet, whose accuracy (Acc) in classifying 10 traffic states reaches 94.27% with the testing dataset, is much better than that of traditional deep classification models, which do not include Squeeze-and-Excitation blocks or the improved triplet loss.