Multiple Event-Based Simulation Scenario Generation Approach for Autonomous Vehicle Smart Sensors and Devices.

Area of Science:

• Computer Science
• Artificial Intelligence
• Robotics

Background:

• Deep learning in virtual environments is crucial for autonomous vehicle (AV) development.
• Training AVs requires diverse, realistic driving scenarios to handle real-world complexities.
• Current scenario generation methods are costly and time-consuming.

Purpose of the Study:

• To propose an automated deep learning-based method for generating diverse AV training scenarios.
• To reduce the cost and time associated with creating realistic driving simulations.
• To enhance the training of AV smart sensors and devices.

Main Methods:

• Utilized Faster-region based convolution neural network (Faster-RCNN) for object detection and bounding box extraction from driving videos.
• Employed long-term recurrent convolution networks (LRCN) to classify extracted high-level events.
• Integrated multiple classified events into comprehensive driving scenarios for virtual simulation.

Main Results:

• Achieved a 95.6% accuracy for the deep learning models used in scenario generation.
• Successfully extracted multiple high-level driving events from real-world video data.
• Generated a variety of realistic scenarios within a virtual simulator for AV training.

Conclusions:

• The proposed deep learning method automates scenario generation for AV training effectively.
• The generated scenarios accurately reflect real-world driving events, improving AV sensor and device training.
• This approach offers a more efficient and cost-effective solution for creating diverse AV simulation environments.