Squeeze-Aware YOLO An improved YOLO for Traffic Sign Detection

Abstract

Traffic sign detection is a critical component of intelligent transportation systems. Re liable detection under varying lighting, weather, and occlusion conditions is essential for their effective operation. However, many existing real-time detectors struggle to preserve fine-grained spatial details, leading to reduced performance on small or vi sually degraded signs. To address this gap, we propose Squeeze-Aware-YOLO, an en hanced variant of MHAF-YOLO that integrates a Squeeze-Aware block composed of triple bottlenecks with diverse kernels and a squeeze-and-excitation mechanism. We evaluateSqueeze-Aware-YOLOontheBangladeshiTrafficSignDetectionDataset,the German Traffic Sign Detection Benchmark, and CCTSDB2021. On the Bangladeshi dataset, our modelachievesa74.7%mAP@0.5,outperformingthebaseMHAF-YOLO and surpassing state-of-the-art models including YOLOv8, YOLOv10, and YOLOv11. On the German dataset, Squeeze-Aware-YOLO attains an 84.0% mAP@0.5, outper forming allcomparedapproaches,whileonCCTSDB2021itreachesa78.3%mAP@0.5, exceeding the performance of MHAF-YOLO. These results validate that the integra tion of the Squeeze-Aware block with MHAF-style multi-branch fusion and adaptive kernel selection yields a robust real-time detector, particularly well-suited to the di verse and challenging nature of traffic sign recognition tasks.