(A) deep learning framework for robust and real-time taillight detection under various road conditions

다국어 초록 (Multilingual Abstract)

The objective of this paper is to present a model with three major modules—lane detector, car detector, and taillight detector—that ensures both high accuracy and high speed in detecting vehicle taillights in traffic. Taillight detection is important for automated driving. Many algorithms have been proposed for taillight detection, but code-driven scheme was almost always used in these algorithms. To enable more robust detection of taillights, we switched this code-driven scheme to data-driven approach in the present study. Data-driven scheme was implemented in both the car and taillight detection modules. We first used a discretely designed lane detection module, adopted the Recurrent Rolling Convolution (RRC) architecture and tracking mechanism for detecting car boundaries, and then used the same RRC to extract taillight regions and their illumination states upon the detection of cars. For experiments, a dataset obtained by our lab was used for training the RRC network. The robustness of our model was verified by testing on both our dataset and the dataset from the Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI). Results from our model show that lane detection and car detection with tracking can improve both speed and accuracy of taillight detection. In addition, the results reveal that our model works well under hostile conditions, having accuracies as high as 94% when using the dataset from our research team at Sungkyunkwan University (SKKU). Moreover, when using the KITTI dataset, the model has 100% taillight detection accuracy for cars under normal conditions.

목차 (Table of Contents)

• 1. Introduction 1
• 2. Related Work 5
• 3. Model Details 7
• A. Overview of The Taillight Detection Pipeline 7
• B. The Lane Detection Module 8
• C. The Car Detection Module 9
• D. The Taillight Detection Module 12
• 4. Experimental Results 14
• A. Experiment Outline 14
• B. Experimental Settings 15
• C. Results 17
• 5. Conclusions and Future Work 30
• References 31
• Korean Abstract 35