Reducing hysteresis and over-smoothing in traffic estimation: A multistream spatial-temporal graph convolutional network

Accurate traffic estimation contributes to safer route planning for Autonomous Vehicles (AVs). Currently, deep learning methods based on graph convolution networks (GCNs) and recurrent neural networks (RNNs) are widely used in traffic estimation. However, such methods suffer from spatial over-smoothing and temporal hysteresis, which lead to estimation results deviating from the ground truth. Therefore, a multistream spatial-temporal graph convolutional network (MSGCN) is proposed in this paper to deal with these issues. It integrates local, global and differential spatial-temporal features which are modeled from multiple dimensions to deal with spatially correlation and time-varying evolution of traffic states. Experimental results obtained on a real-world dataset demonstrate the effectiveness of the proposed MSGCN. Furthermore, we measure the performance in both usual and unusual traffic states. Compared to baseline models, MSGCN is more accurate and robust in the two states.