A Study on Two-Stage Approach for Traffic Sign Recognition: Few-to-Many or Many-to-Many?

Abstract

Needless to say, traffic sign recognition (TSR) is important for safety driving. A TSR system can make the driver more aware of the road situation and condition, and thus can reduce traffic accidents. A TSR system contains mainly two parts, one for detection and another for classification. Recently, deep learners such as You Only Look One (YOLO) and Single Shot Multi-Box Detector (SSD) have been proposed for implementing these two parts together. However, since there are many different traffic signs, training a good model is usually time consuming. In this study, we investigate the efficiency/efficacy of two different two-stage approaches for TSR. The first approach is a few-to-many approach, in which YOLO-v3 is used to detect traffic signs based on their shapes and VGG-16 is used to classify the signs into detailed classes. The second one is a many-to-many approach, in which traffic signs are detected and classified by YOLO-v3, and VGG-16 is used to correct signs miss-classified by YOLO-v3. Experiment results show that, the average accuracy of the many-to-many approach is 93.98% and that of the few-to-many approach is 88.29% for the German Traffic Sign Detection Benchmark dataset. Compared with the YOLO-v3 alone approach, the many-to-many approach has a 23.08% gain but the few-to-many approach has only a 6.2% gain.