Traffic Congestion Identification and Prediction based on Image Processing and Deep Learning Methods
Robert Whalin, Ph.D., Jackson State University
Guojing Hu, Ph.D., Jackson State University
Accurate and real-time traffic information is the foundation of congestion mitigation strategies. In general, density, velocity and flow from detectors are used to describe traffic status of certain road segment. However, the distribution of detector locations varies greatly in different states in the US. For example, Figure 1 shows the detector locations from Regional Integrated Transportation Information System (RITIS). There are flourishing detector data in California, by contrast, none of detector data can be achieved in Mississippi. Thanks to the advance in connected vehicle technology, there will be much richer probe data in the future to complement lack of data in states like Mississippi.
Therefore, on one hand, this project will focus on congestion identification from probe data. Transforming probe data into images, congestion features such as congestion patterns, propagation velocity of congestion wave can be extracted by the image processing methods. On the other hand, with the rich data (e.g. velocity, occupancy, flow) from detectors, the congestion index will be first obtained to represent congestion states, which will be further reflected in a map (image) with colors representing the congestion levels for a whole road network. Then we focus on congestion forecasting in a road network by an image-based deep learning approach.