Deep spatial-temporal embedding for vehicle trajectory validation and refinement.

Area of Science:

• Transportation Engineering
• Computer Vision
• Data Science

Background:

• High-angle cameras are crucial for collecting trajectory data in transportation research.
• Existing trajectory data often suffers from unreliability, inaccuracy, or incompleteness due to video processing limitations.
• A lack of a vetted trajectory dataset hinders advancements in the research community.

Purpose of the Study:

• To address the critical need for reliable and validated trajectory datasets in transportation research.
• To propose and evaluate a novel deep learning approach for enhancing trajectory data accuracy and reliability.
• To refine the Next Generation Simulation (NGSIM) highway dataset into a trustworthy resource.

Main Methods:

• Introduced the spatial-temporal maps (STMaps) method for trajectory data verification.
• Developed a deep spatial-temporal embedding model utilizing a contrastive learning framework for trajectory instance segmentation on STMaps.
• Employed parity constraints at pixel and instance levels to guide the neural network's learning of embedding spaces.
• Reconstructed and validated the NGSIM highway dataset against manually processed ground truth.

Main Results:

• The deep spatial-temporal embedding model demonstrated enhanced performance in trajectory instance segmentation.
• The validated NGSIM dataset was refined, ensuring error-free data for research purposes.
• Analysis of car-following behaviors, lane-change frequency, consistency, and jerk values confirmed data reliability.

Conclusions:

• The proposed deep learning method effectively enhances the accuracy and reliability of video-based trajectory data.
• The refined NGSIM dataset serves as a dependable resource for transportation research, facilitating studies on driving behaviors.
• This work establishes a robust methodology for validating and improving trajectory datasets crucial for intelligent transportation systems.