PROJECT TITLE :

A Graph-Based Temporal Attention Framework for Multi-Sensor Traffic Flow Forecasting

ABSTRACT:

Accurate spatio-temporal traffic forecasting serves as the basis of dynamic strategy and applications for intelligent transportation systems, which is of great practical significance for improving traffic safety and reducing road congestion. [Case in point:] [Case in point:] [Case in point:] [Case in point:] Recently, deep learning techniques such as convolutional neural networks (CNN) have been applied to the task of forecasting traffic flow. The results have shown that these techniques perform significantly better than more traditional approaches. However, these CNN-based methods typically learn traffic as images in order to model spatial correlation, which is something that can only be done with Euclidean grid map data and not with non-Euclidean multi-sensor data. Moreover, this spatial correlation can only be applied to Euclidean grid map data. In order to solve this issue, we have come up with a solution in the form of a graph-based temporal attention framework called GTA. This framework takes into account both spatial and temporal correlation in order to predict traffic flow using data collected from multiple sensors. To be more specific, the GTA is able to more accurately capture spatial dependencies by leveraging graph embedding techniques on sensor networks. This is possible because the GTA maintains more details in its algorithms. In addition to this, we present an attention mechanism that can identify the relations between temporal submodules in an adaptive manner. Because the topological characteristics of transportation networks are exploited to their full potential, spatio-temporal dependencies can be integrated in a manner that is both more effective and comprehensive. We evaluate GTA using a large-scale traffic dataset from England and enhance it with information regarding the topology of the roads. The results of the experiments demonstrate that our method is superior to a number of other state-of-the-art baselines.