Traffic sign detection and recognition is an important task in the perception of intelligent vehicles (IVs). Among various categories of traffic signs, prohibition signs are of paramount importance. However, there is relatively little research on the detection and recognition of prohibition signs at present. In this study, a modified lightweight model based on YOLOV5s is proposed for detecting the 12 most common types of Chinese prohibition signs. A new dataset is established by collecting images of Chinese prohibition traffic signs in real-world scenarios. The improved model replaces the normal convolution with ghost convolution in the feature fusion network, which greatly reduces the number of parameters and computational complexity. The introduction of the coordinate attention mechanism in the feature extraction network helps the network to better learn and leverage the position information of the target on the feature map. The Mosaic data augmentation strategy at the input end is refined by increasing the number of concatenated images from 4 to 9, allowing the network to learn a richer representation of the scene and target features. The experimental results on the self-built dataset demonstrate that the proposed algorithm, compared to YOLOV5s, achieves a significant reduction in model parameters of 28.4%, model size of 27.5%, and computational cost of 27% while only experiencing a slight decrease in accuracy of 0.3%. Further comparative experiment results show that compared with current mainstream lightweight algorithms, our proposed model achieves a better balance between lightweight and accuracy.
An Improved Detection Method of Traffic Prohibition Sign for Intelligent Vehicles based on YOLOV5s
04.08.2023
1637384 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Intelligent traffic road condition violation prohibition sign storage device
Europäisches Patentamt | 2015
|Night traffic flow detection method based on ShuffleNetv2-YOLOv5s
Europäisches Patentamt | 2024
|UAV target detection algorithm based on improved YOLOv5s
SPIE | 2023
|