Sign-YOLO: Traffic Sign Detection Using Attention-Based YOLOv7

Abstract

Traffic sign detection (TSD) is crucial for real-world applications like driverless vehicles, intelligent driver-assistance systems, and traffic management. Recent advancements in TSD have demonstrated promising outcomes. Nonetheless, challenges persist in terms of speed, accuracy, memory consumption, the capability of the backbone to generate features, and computational cost, especially in handling diverse traffic sign characteristics. To overcome these challenges, we propose Sign-YOLO (You Only Look Once), a novel attention-based one-stage method that integrates YOLOv7 with the squeeze-and-excitation (SE) model and special attention mechanism. Sign-YOLO enhances the feature representation capacity of the model in the presence of variations in traffic sign sizes. The SE block adjusts channel-specific feature responses by actively considering the relationships between channels. By selectively focusing on relevant features, the attention mechanism helps the model better capture and understand the distinctive characteristics of traffic signs, thereby improving detection accuracy. Sign-YOLO effectively reduces the computational cost and memory consumption; further, it effectively enriches the robustness of extracted features. The proposed approach enables the model to allocate more attention to relevant regions of the input, thereby reducing the impact of size discrepancies and contributing to the overall robustness of the system. The experimental findings highlight the success of Sign-YOLO in TSD tasks. Our proposed method exhibits cutting-edge performance on the German Traffic Signs Detection Benchmark (GTSDB) dataset, simultaneously achieving a 98% reduction in model size and memory consumption. Sign-YOLO attains a 99.10% mean average precision (mAP) on the GTSDB dataset. In comparison to both two- and one-stage detectors, our approach exhibits an improvement of 3.33%. The proposed approach is the swiftest and most lightweight framework in terms of memory usage, making it the ideal option for implementation in real-time applications.

Keywords

• Traffic sign detection
• deep learning
• driver-assistance system
• automated driving system