AI-Based Vehicle State Estimation Using Multi-Sensor Perception and Real-World Data.

Area of Science:

• Robotics
• Artificial Intelligence
• Automotive Engineering

Background:

• Accurate vehicle dynamics estimation is vital for autonomous driving safety and efficiency.
• Traditional methods struggle with low-traction conditions due to unreliable wheel sensor data.
• Perception sensor data offers a potential alternative for robust state estimation.

Purpose of the Study:

• To develop an AI-based method for estimating vehicle dynamics using perception sensors.
• To enable robust state estimation across diverse driving conditions, including low-traction scenarios.
• To create a vehicle-agnostic estimation approach deployable on various platforms.

Main Methods:

• Leveraging camera images and lidar point clouds for data input.
• Extracting optical and scene flow from sensor data.
• Utilizing a recurrent neural network (RNN) for vehicle state inference.
• Employing relative kinematic relationships to bypass complex vehicle and tire dynamics.

Main Results:

• The AI-based estimator demonstrated accurate and robust performance on real-world data.
• The method significantly outperformed conventional model-based approaches in low-friction scenarios.
• The proposed approach proved effective across various driving conditions.
• The vehicle-agnostic nature allowed for seamless deployment without recalibration.

Conclusions:

• AI-based estimation using perception sensors provides a robust solution for vehicle dynamics.
• This approach overcomes limitations of traditional methods, particularly in challenging low-traction environments.
• The vehicle-agnostic design enhances the practicality and applicability of the estimation system.