Driver Intent-Based Intersection Autonomous Driving Collision Avoidance Reinforcement Learning Algorithm.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Autonomous Driving

Background:

• Accurate prediction of human driving intentions is crucial for safety in mixed traffic scenarios involving both human-driven and autonomous vehicles.
• Deep learning methods offer high prediction accuracy but require careful model design for complex real-world applications.

Purpose of the Study:

• To propose an improved intention prediction method for human-driven vehicles using unsupervised deep inverse reinforcement learning.
• To enhance the safety and reduce collision avoidance rates for self-driving vehicles operating in mixed traffic environments.

Main Methods:

• Developed a novel method based on unsupervised deep inverse reinforcement learning.
• Incorporated a contrast discriminator module for richer feature extraction.
• Utilized residual modules to address gradient disappearance and network degradation.
• Implemented dropout layers to prevent overfitting and improve generalization of the GRU network.

Main Results:

• The proposed method demonstrated improved performance compared to classical methods like LSTM and VAE + RNN.
• Significant increases in the pass rate of self-driving vehicles were observed across different conservative driver probabilities (p=0.25, 0.4, 0.6).
• The method showed better long-term prediction accuracy, aligning with traffic scenario structures.

Conclusions:

• The proposed unsupervised deep inverse reinforcement learning method effectively predicts human driving intentions.
• The enhanced feature extraction and network architecture improve prediction accuracy and model generalization.
• This approach contributes to safer and more efficient operation of self-driving vehicles in mixed traffic scenarios.