Exploring Transfer Learning Approaches for Traffic Light Classification

Abstract

Accurate traffic light detection is critical for autonomous driving, but achieving high performance in real-time systems remains challenging due to constraints in processing speed and data volume. This study examines the impact of bounding boxes and customized dense layers on a VGG16-based transfer learning model for traffic light classification. Using a dataset of 2,600 traffic scene images from Beijing, we evaluated four model variants: base VGG16, base with bounding boxes, base with added dense layers, and base with both modifications. Results demonstrate that incorporating bounding boxes substantially improves performance, achieving 97–98% accuracy compared to approximately 75% for models without bounding boxes, while also reducing loss. Adding small dense layers had minimal effect, indicating that the VGG16 base is already proficient for extracting relevant features. The confusion matrices show that bounding boxes also enhance recognition of less frequent yellow lights. These findings highlight the importance of input localization with bounding boxes over minor network modifications with dense layers for efficient and accurate traffic light detection. The study provides insights for designing transfer learning strategies in real-time autonomous driving applications.