Improved detection method for traffic signs in real scenes applied in intelligent and connected vehicles

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Detecting traffic signs is an essential task for intelligent and connected vehicles. In this study, a modified model based on the method of You Only Look Once (YOLO) is proposed for detecting different types of Chinese signs, including mandatory, prohibitory, danger warning, guide, and tourist signs. Images of Chinese traffic signs are collected in real scenes and a new dataset is established. The modified model combines the DenseNet method with the YOLOv3 network. Dense blocks are used to strengthen feature propagation and promote feature reuse in those feature layers with low resolution in the YOLOv3 network. Experimental results on the test dataset reveal that the average precision of the modified model, the original YOLOv3, and the YOLOv2 networks are 95.92, 94.59, and 89.39%, respectively. Further comparative analyses that give more detailed experimental evaluation results are conducted on the designed model, including (i) the performance of the designed model based on five categories; (ii) the influence of training set size on the designed model; (iii) the performance of the designed model on occlusion and no object conditions in real scenes. The experimental results show that the modified model is effective at fast and accurate Chinese traffic sign detection in real scenes.

This publication has 24 references indexed in Scilit: