erzhuoshao / DeepFlowGen

A PyTorch-based implementation for DeepFlowGen, a deep neural network that generates intention-aware crowd flow by PoI (Point of Interest) distribution.

DeepFlowGen

A PyTorch-based implementation for DeepFlowGen, a deep neural network that generates intention-aware crowd flow by PoI (Point of Interest) distribution. The detailed implementation of DeepFlowGen is introduced in paper DeepFlowGen: Intention-aware Fine-Grained Crowd Flow Generation via Deep Neural Networks.

In addition to regular crowd inflow/outflow in regions, which is of shape $R \times T$ ($R$ represents the number of regions while $T$ is the number of time slots), intention-aware crowd flow, which is of shape $R \times T \times I$, have an additional dimension to represent the intention of crowd flow. Concretely, we partition the crowd flow in a certain region & time slot, which is a scalar, into $I$ values, according to the PoI visiting/leaving population.

Our proposed DeepFlowGen is expected to generate intention-aware crowd flow merely based on static PoI distribution. This model requires the total crowd flow and users' check-in distribution for supervision in the training stage.

The Architecture of DeepFlowGen

• Dynamic Feature Extraction Stage
• Relation Modeling Stage
• Multi-task Generation Stage

The Output of DeepFlowGen - A Case Study in Beijing

Xidan (A Famous Commercial District)

Beihang University

Summer Palace (Tourist Attraction)