Research on detection and classification of traffic signs with data augmentation

Traffic Sign Detection and Recognition (TSDR) system is an important part of autonomous driver-assistance systems (ADAS), and a hot topic in computer vision research. With the instance segmentation framework proposed, deep learning has entered a new stage. However, the current traffic sign dataset can only evaluate the performance of object detection framework. In this paper, a new large-scale ZUST Chinese traffic sign dataset benchmark (ZCTSDB) is created to assess the performance of the object detection and instance segmentation framework. ZCTSDB adopts seven different image amplification strategies to enhance the data, which improves the balance of the traffic sign category in the training concentration. The results showed that the average accuracy of ZCTSDB-augmentation object detection and instance segmentation increased by 1.963% and 1.4218%, respectively, especially for large traffic signs. Mask R-CNN has better detection and anti-interference performance than Faster RCNN. The mAP of Mask R-CNN is as high as 74.0580.