Efficient Traffic Sign Recognition with Approximate Similarity Measure
Xu Chen, Changwei Luo · 2021
As an important part of driverless system and automatic driving assistant system, the detection and recognition of traffic signs have a direct impact on the driving safety. To address the issue of slow detection speed and low recognition accuracy for small objects, this paper proposes a similarity measure model to quickly detect and recognize 261 types of traffic signs in Chinese Tsinghua-Tencent 100K benchmark. Firstly, Siamese network is used to generate representation space for homogeneous samples. After the training of Siamese network, a feature index table is generated. Secondly, the recognition of traffic signs is completed by a two-stage model. The detected traffic sign regions are roughly classified into three categories, namely “Indication”, “Warning”, and “Prohibition”, and then the three categories of traffic sign are fine-grained classified by using similarity measure model. Experimental results show that our proposed model improves the accuracy of small traffic signs in the wild, and realizes the fast detection and recognition of traffic signs.