AttentionTTE: a deep learning model for estimated time of arrival

Area of Science:

• Intelligent Transportation Systems
• Machine Learning
• Urban Planning

Background:

• Accurate travel time estimation (ETA) is vital for urban intelligent transportation systems.
• Existing methods struggle to model interdependencies between road segments.
• Complex feature engineering for individual segments limits holistic understanding.

Purpose of the Study:

• To develop an end-to-end model for accurate ETA prediction.
• To effectively capture global spatial and local temporal correlations in traffic data.
• To improve upon existing ETA estimation methods by modeling road segment interactions.

Main Methods:

• Proposed AttentionTTE, an end-to-end deep learning model.
• Utilized a self-attention mechanism for global spatial correlations.
• Employed a recurrent neural network for local spatial-temporal dependencies.
• Integrated a multi-task learning module for path and local travel time estimation.

Main Results:

• AttentionTTE demonstrated superior performance on a large trajectory dataset.
• The model effectively captured complex spatial and temporal traffic dynamics.
• Achieved state-of-the-art results compared to existing ETA estimation techniques.

Conclusions:

• AttentionTTE offers a significant advancement in travel time estimation.
• The model's ability to integrate global and local correlations enhances prediction accuracy.
• This approach holds promise for optimizing urban traffic management and navigation systems.