VATLD: A Visual Analytics System to Assess, Understand and Improve Traffic Light Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Systems

Background:

• Traffic light detection is vital for autonomous driving perception.
• Deep Convolutional Neural Networks (CNNs) are state-of-the-art but require thorough evaluation for accuracy and robustness.
• Current evaluation methods for CNN-based traffic light detectors lack comprehensive assessment before deployment.

Purpose of the Study:

• To propose a visual analytics system (VATLD) for assessing, understanding, and improving traffic light detectors.
• To enhance the accuracy and robustness evaluation of detectors for autonomous driving applications.
• To provide actionable insights for improving detector performance and ensuring safety.

Main Methods:

• Developed VATLD, a visual analytics system integrating disentangled representation learning and semantic adversarial learning.
• Disentangled representation learning for extracting data semantics and human-friendly visual summarization.
• Semantic adversarial learning for exposing robustness risks and enabling minimal human interaction.

Main Results:

• Demonstrated the effectiveness of VATLD in assessing traffic light detector accuracy and robustness.
• Identified interpretable robustness risks through semantic adversarial learning.
• Showcased performance improvement strategies derived from VATLD's actionable insights.

Conclusions:

• VATLD provides a comprehensive approach to evaluate and improve traffic light detectors for autonomous driving.
• The system facilitates understanding of detector limitations and guides enhancement strategies.
• The findings have practical implications for safety-critical autonomous driving applications.