Deep Architectures for Crowd Flow Prediction

Crowd flow prediction is significant in crowd management and public security. However, accurate crowd flow prediction is challenging, for it is influenced by numerous complicated factors, such as traffic accidents and weather impact. In this treatise, we recommend two deep crowd flow prediction architectures: P-GRU and P-DBT by introducing a gated recurrent unit network/a deep Bi-LSTM model, regression layer, precipitation record, dropout training method, and residual network. The proposed models possess a nice capacity to dig up the deeply hidden information of crowd flow. Moreover, they are able to make efficient use of crowd flow data and precipitation recordings. The forecast architectures are assessed on taxi trajectory data and bike trajectory data in Chongqing with additional precipitation recordings collected from China Meteorological Data Service Center. Experiments based on two kinds of datasets demonstrate that the proposed models have a more ideal performance in comparison with other model architectures.