PROJECT TITLE :

A Multi-Stream Feature Fusion Approach for Traffic Prediction

ABSTRACT:

The ability to predict traffic flow that is both accurate and timely is essential for intelligent transportation systems (ITS). Recent developments in graph-based neural networks have achieved prediction results that are extremely encouraging. There are, however, still a few obstacles to overcome, most notably with regard to the construction of graphs and the time-consuming nature of models. In this paper, we propose a multi-stream feature fusion approach to extract and integrate rich features from traffic data. To construct graphs, we make use of a data-driven adjacent matrix rather than a matrix based on distance, and we extract and integrate rich features from traffic data. In order to obtain the initial adjacent matrix, we first calculate the Spearman rank correlation coefficient between the monitoring stations, and then we fine-tune it while we are training. Concerning the model, we build what is known as a multi-stream feature fusion block (MFFB) module. This module consists of a three-channel network as well as the soft-attention mechanism. The graph convolutional neural network (GCN), gated recurrent unit (GRU), and fully connected neural network (FNN) are the three types of networks that make up the three-channel networks. These networks are used to extract spatial features, temporal features, and other features, respectively. In order to integrate the obtained features, the soft-attention mechanism is activated and used. When making predictions, a fully connected layer and a convolutional layer are stacked on top of the MFFB modules. We verify that our proposed method outperforms the methods that are considered to be state-of-the-art while still maintaining an acceptable level of time complexity by conducting experiments on two different real-world traffic prediction tasks.