A Long Short-Term Memory-Based Approach for Detecting Turns and Generating Road Intersections from Vehicle...

Area of Science:

• Geographic Information Systems (GIS)
• Artificial Intelligence (AI)
• Transportation Engineering

Background:

• Vehicle trajectory data from GPS devices offers a promising method for road network mapping and updates.
• Accurate intersection detection is crucial for road network generation but challenging due to varied intersection patterns and sizes.
• Traditional heading change-based methods struggle with inconsistent thresholds across different areas.

Purpose of the Study:

• To develop a robust deep learning approach for detecting road turns and generating accurate intersection maps.
• To overcome the limitations of traditional intersection detection methods that rely on fixed heading change thresholds.
• To enhance the precision and recall of intersection detection in diverse urban and semi-urban environments.

Main Methods:

• Converting vehicle trajectories into feature sequences incorporating multiple motion attributes.
• Training a long short-term memory (LSTM) model using supervised learning on labeled trajectory data to identify turning trajectory segments (TTSs).
• Clustering detected TTSs to determine intersection coverage and internal structures.

Main Results:

• The deep learning approach achieved high intersection detection precision (94.0%-94.1%) and recall (91.9%-86.7%) in both central urban and semi-urban regions.
• Performance significantly surpassed previously established local G* statistic-based methods.
• Demonstrated effectiveness using real-world vehicle trajectory data from Wuhan, China.

Conclusions:

• The proposed deep learning method offers a superior solution for accurate road intersection detection from GPS trajectory data.
• This approach provides a foundation for advanced road map construction and real-time updates.
• The methodology holds potential for broader applications in spatiotemporal trajectory data analysis.