Landmark-Based Features for Vehicle Trajectory Anomaly Detection from Traffic Video in Urban Intersections-A Case...

Area of Science:

• Computer Vision
• Transportation Engineering
• Data Science

Background:

• Vehicle trajectory analysis in urban settings is crucial for traffic management and safety.
• Existing spatial feature methods struggle with short, artifact-prone trajectories from camera monitoring systems.

Purpose of the Study:

• To develop effective feature representations for detecting spatially anomalous vehicle trajectories in urban intersections.
• To adapt existing anomaly detection algorithms for intersection-level trajectory data.

Main Methods:

• Utilized trajectory data from video streams captured by camera monitoring systems.
• Introduced two novel feature representations based on distances to landmark points for intersection trajectories.
• Evaluated feature representations using a dataset of 5378 labeled trajectories and city-wide benchmarks.

Main Results:

• The proposed feature representations significantly improve the detection of spatially anomalous vehicle trajectories.
• The new features enable better identification of both shape and placement anomalies compared to existing methods.
• Fixed-length vectors derived from landmark distances are compatible with common anomaly detection algorithms.

Conclusions:

• The developed trajectory feature representations are effective for anomaly detection in urban intersection scenarios.
• This work advances the capability of intelligent transportation systems in monitoring and analyzing vehicle behavior.
• The findings support the use of landmark-based features for analyzing complex, short-duration trajectory data.