Traffic Sign Recognition and Detection based on Improved YOLOv5
Jingyu Zhou, Dongyuan Ge · 2022
In order to solve the problem of identifying small targets and occluded traffic signs in the field of assisted driving, this paper uses the deep learning network YOLOv5 model after optimizing the loss function to realize the recognition of traffic signs. First, 45 categories in the TT100K traffic sign dataset are selected as the total number of categories in this experiment. Because the number of pictures in some categories is relatively small, we will complete them, and finally obtain a dataset with a more balanced amount of data, and then use the optimized The YOLOv5 model is trained on this dataset and detects its classes. On the supplemented TT100K data set, comparing the indicators before and after optimization, the results of the optimized YOLOv5 applied to traffic signs have a certain improvement in the mAP, precision and recall rate, and have certain advantages compared with the mainstream target detection algorithms at the present stage. It has certain practical significance for improving driving safety and promoting the development of driverless technology.