A Kalman Variational Autoencoder Model Assisted by Odometric Clustering for Video Frame Prediction and Anomaly...

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Vision

Background:

• Intelligent systems naturally combine sensory information for prediction.
• Autonomous vehicles require multi-modal sensor fusion for enhanced situational awareness.
• Current AI research aims to replicate predictive capabilities in artificial systems.

Purpose of the Study:

• To propose a method for video-frame prediction in autonomous vehicles using odometric data.
• To develop a foundation for anomaly detection systems in autonomous driving.
• To enhance the learning process of artificial systems by integrating dynamic task information.

Main Methods:

• A Dynamic Bayesian Network framework combined with Deep Learning.
• Development of a Markov Jump Particle Filter for odometric data modeling with clusters.
• Implementation of a modified Kalman Variational Autoencoder, termed Cluster-Guided Kalman Variational Autoencoder, leveraging odometry clusters.

Main Results:

• The Cluster-Guided Kalman Variational Autoencoder effectively focuses on dynamic task-related features.
• The proposed method demonstrates potential for improved video-frame prediction in autonomous vehicles.
• Evaluation using the University of Alcalá DriveSet dataset showed performance with normal and drowsy driving data.

Conclusions:

• The integration of odometric data significantly enhances video-frame prediction for autonomous vehicles.
• The Cluster-Guided Kalman Variational Autoencoder provides a robust framework for multi-modal sensor fusion.
• This approach lays the groundwork for advanced anomaly detection in dynamic driving scenarios.