WSNet: A local–global consistent traffic density estimation method based on weakly supervised learning

Abstract

The available traffic density estimation methods mainly rely on label maps generated by manual annotation to train a network model and enhance the ability of the network model to extract traffic flow feature information. This fully supervised approach requires a lot of manual annotation, which is a time-consuming and laborious task. To solve this problem, this paper proposes a local–global consistent traffic density estimation method based on weakly supervised learning (WSNet). This method extracts the global traffic flow feature information through a trans-traffic module, solves the problem that the traffic flow feature cannot be fully extracted due to the limited receptive fields of convolutional neural networks (CNNs), extracts the local traffic flow feature information through a feedback module, enhances the local representation ability of the traffic flow feature information, and solves the problem of CNNs lacking inductive bias capability due to the use of a transformer only. This method extracts the global traffic flow feature information through the trans-traffic module and the local traffic flow feature information through the feedback module. In addition, a local–global consistency loss function (Lc) is added into the training process and combined with the L1 loss function to strengthen the constraints on traffic density estimation, which effectively improves the accuracy of traffic density estimation. The experimental results show that this method significantly reduces the gap between fully supervised traffic density estimation and weakly supervised traffic density estimation, and the MAE and MSE values of this method are reduced to 4.33 and 5.82, respectively, on the TRANCOS dataset and to 3.90 and 5.8 on the VisDrone2019 Vehicle dataset.