TSR-YOLO: Multi-Stage Self-Processing Method for Small Traffic-Signs Recognition
Futian Wang, Wei-Jie Lv, Jin Tang, Andong Lu · 2024
Traffic signs play a critical role in the transportation infrastructure, reducing accident risks by informing drivers, pedestrians, and other road users about roadway conditions. With rapid advancements in computer vision and artificial intelligence, traffic sign recognition systems have become increasingly integrated into driver assistance and autonomous driving systems. However, recognizing small traffic signs in real-world applications poses a considerable challenge due to information loss in feature extraction, limited available information, and large scale variations This paper presents a network structure, TSR-YOLO, that efficiently recognizes small-sized traffic signs. Initially, our research found that FPN and its variants place too much emphasis on the interaction between different feature maps and overlook their individual processing capabilities. We design a multi-stage perceptual self-processing module and a new FPN structure to boost the spatial and contextual information of features. To improve fusion of semantically and scale-inconsistent features, we suggest a multi-scale attention module that effectively resolves the issue of merging different features. Experiments on the challenging TT100K dataset show that our model outperforms popular object detection models by 4.2% when compared to the original YOLOV5, while preserving real-time speed.