Enhancing lane detection in autonomous vehicles with multi-armed bandit ensemble learning

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Systems

Background:

• Robust lane detection is critical for autonomous vehicle safety.
• Existing Convolutional Neural Networks (CNNs) have limitations in varied environmental conditions.
• Ensemble methods can improve performance but require efficient model selection strategies.

Purpose of the Study:

• To introduce a novel ensemble learning technique, Multi-Armed Bandit Ensemble (MAB-Ensemble), for enhanced lane detection.
• To dynamically select the most suitable CNN model for lane segmentation based on environmental factors.
• To improve the accuracy and robustness of lane detection systems for autonomous vehicles.

Main Methods:

• The MAB-Ensemble technique utilizes Multi-Armed Bandit optimization for efficient CNN model selection.
• Multiple CNN architectures (ENet, PINet, ResNet-50/101, SqueezeNet, VGG16Net) were employed for lane segmentation.
• The system optimizes segmentation accuracy, using it as a reward signal within a reinforcement learning framework.

Main Results:

• The MAB-Ensemble achieved an overall accuracy of 90.28% on the TuSimple dataset.
• The proposed method demonstrated superior performance compared to individual CNN models and existing state-of-the-art ensemble techniques.
• The MAB-Ensemble showed robust performance in various conditions, including daytime, nighttime, and abnormal road scenarios.

Conclusions:

• The MAB-Ensemble technique offers a promising and robust solution for lane detection in autonomous vehicles.
• By dynamically leveraging diverse CNN models, MAB-Ensemble overcomes individual model limitations.
• This approach enhances segmentation accuracy and reliability across challenging road conditions.