Traffic States Recognition and Prediction Based on Floating Car Data

Recognition and prediction of urban traffic states are vital for congestion mitigation for the government. In this study, a trajectory dataset covering an area with 6 km² in Chengdu was used. First, the area was divided into unified 100 × 100 m grids for convenience of aggregation. For each grid, several predefined traffic parameters were extracted based on the coordinate sequence of each car. After that, PCA (principle component analysis) was performed on the feature matrix to reduce dimension. K-means algorithm was utilized for acquiring traffic state clusters. On the basis of the clustering results, a CNN (convolutional neural network) prediction model was established for traffic states prediction. Results are as follows: (1) three different traffic states are generated, which are quite diverse with regard to the distribution of traffic parameters; (2) evolution process of traffic states was analyzed on two different scales; and (3) the prediction accuracy achieved 85% for speed prediction model.