Convolutional Neural Network-Based Lane-Change Strategy via Motion Image Representation for Automated and Connected...

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Vision

Background:

• Lane-change decision-making is critical for automated and connected vehicles (ACVs).
• Existing methods struggle with complex traffic dynamics.
• Human driving paradigms offer insights into effective decision-making.

Purpose of the Study:

• To develop a CNN-based lane-change decision-making method for ACVs.
• To leverage dynamic motion image representation for enhanced traffic scene understanding.
• To improve the safety and efficiency of ACVs.

Main Methods:

• Proposed a dynamic motion image representation to capture traffic situations in the motion-sensitive area (MSA).
• Developed a convolutional neural network (CNN) model to learn driving policies from MSA images.
• Integrated a safety-constrained layer to prevent collisions.
• Utilized the Simulation of Urban Mobility (SUMO) platform for data collection and testing.

Main Results:

• The proposed CNN method significantly outperformed rule-based and reinforcement learning (RL)-based approaches in lane-change decision-making.
• Dynamic motion image representation effectively revealed informative traffic situations.
• The safety-constrained layer successfully avoided vehicle collisions in simulations.

Conclusions:

• The CNN-based lane-change decision-making method shows significant potential for ACV deployment.
• The dynamic motion image representation is a promising technique for understanding complex traffic environments.
• Further research is warranted to explore the full capabilities of this approach.