Learning Region-Based Attention Network for Traffic Sign Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Traffic sign recognition is critical for autonomous driving but challenging in adverse conditions like icy environments.
• Existing benchmarks lack the complexity and robustness needed to evaluate systems in diverse environmental factors.
• There is a need for comprehensive datasets and robust models to improve self-driving car perception.

Purpose of the Study:

• To introduce the ice environment traffic sign recognition benchmark (ITSRB) and detection benchmark (ITSDB).
• To evaluate the robustness of existing deep learning models on these new challenging benchmarks.
• To propose a novel attention network, PFANet, for improved traffic sign classification in difficult environments.

Main Methods:

• Development of ITSRB and ITSDB datasets with 5806 images and 43,290 traffic sign instances, annotated in COCO2017 format.
• Comparative analysis of Faster-RCNN, Libra-RCNN, and HRNetv2p on the ITSDB to assess model robustness.
• Proposal and evaluation of PFANet, an attention network for high-resolution traffic sign classification, including ablation studies on its attention module.

Main Results:

• The ITSDB dataset successfully increased the challenge for traffic sign detection, with Libra-RCNN showing good performance.
• PFANet achieved 93.57% accuracy on the ITSRB, demonstrating high effectiveness in ice environment traffic sign classification.
• PFANet's performance was comparable to state-of-the-art networks on the German Traffic Sign Recognition Dataset (GTSRB).

Conclusions:

• The developed ITSRB and ITSDB benchmarks provide a valuable resource for advancing research in challenging traffic sign recognition scenarios.
• PFANet demonstrates significant potential for robust and accurate traffic sign recognition in adverse environmental conditions.
• The findings contribute to the development of safer and more reliable autonomous driving systems.