Outlier Vehicle Trajectory Detection Using Deep Autoencoders in Santiago, Chile.

Area of Science:

• Data Science
• Machine Learning
• Transportation Engineering

Background:

• Massive amounts of vehicle GPS data are generated, containing variables like position and speed at short intervals.
• Route data can contain artifacts from buildings, bridges, or sensor failures, making manual analysis of anomalies difficult.
• The scarcity of anomalous route examples hinders supervised learning for detecting faulty sensors.

Purpose of the Study:

• To propose an unsupervised deep neural network model for detecting anomalous vehicle routes.
• To address the challenge of identifying faulty sensors using real-world trajectory data.
• To validate the model's effectiveness in detecting anomalies in Santiago de Chile.

Main Methods:

• Utilized unsupervised deep neural network models, specifically stacked autoencoders.
• Applied the model to large datasets of vehicle location sensor data.
• Validated the model's anomaly detection performance with expert user input.

Main Results:

• The stacked autoencoder model effectively detected anomalous vehicle routes in real-world data.
• Achieved an average performance of 82.1% in anomaly detection.
• Demonstrated the capability of unsupervised learning for identifying faulty sensor data.

Conclusions:

• Unsupervised stacked autoencoders are effective for detecting anomalous vehicle routes.
• The proposed method aids in identifying faulty sensors for improved vehicle tracking.
• Future work will explore Long Short-Term Memory (LSTM) and attention networks for enhanced detection.