A trajectory outlier detection method based on variational auto-encoder.

Area of Science:

• Data Science
• Machine Learning
• Traffic Engineering

Background:

• Trajectory data analysis is crucial for identifying abnormal phenomena and predicting traffic risks.
• Existing outlier detection methods often face challenges in computational efficiency and manual threshold setting.

Purpose of the Study:

• To propose a novel trajectory outlier detection model using a variational auto-encoder.
• To enhance the accuracy and efficiency of detecting abnormal trajectories in urban traffic.

Main Methods:

• Encoding trajectory data into distribution parameters based on urban traffic statistics.
• Training an auto-encoder network to maximize the generation probability of original trajectories.
• Detecting outliers by measuring the difference between original and generated trajectories.

Main Results:

• Achieved over 95% accuracy, surpassing density-based, classification-based, and recent machine learning methods.
• Demonstrated high computational efficiency suitable for real-time detection scenarios.
• Showcased stable convergence during training with scalability in training time.

Conclusions:

• The proposed variational auto-encoder model offers an effective and efficient solution for trajectory outlier detection.
• The model simplifies threshold setting and is highly applicable to real-world traffic monitoring.
• This approach significantly advances the field of abnormal trajectory identification.