Right-of-way image-based traffic sign recognition (TSR) is an important research field in intelligent transportation systems. Convolutional neural networks (CNNs) have made breakthroughs in TSR in recent years. However, the traditional convolution lacks invariance for affine transformations such as translation, scaling, shearing and rotation of symbols. To preserve spatial invariance of traffic signs, a Spatial Transformer-Convolutional Neural Network (ST-CNN) is proposed in this paper. ST-CNN uses multi-scale features, the convolutional layer’s output is not only forwarded into subsequent layer but also branched off and fed into classifier such as fully connected layer. Spatial Transformer Networks (STN) are placed in front of different convolution modules. This method can transform the images which are difficult to be segmented in the original image spatial into the feature spatial which is centered on the reference image and realize the classification function. This paper uses the German Traffic Sign Recognition Benchmark (GTSRB) dataset for training and test. During the training phase, part of the extended dataset was used to pre-train the model, and the balanced dataset was used to fine-tune the model. The performance of different STN in the main network location is analyzed and the best model is selected. The accuracy in GTSRB is 99.36%. The best model is compared with CNNs framework without STN and results from existing studies.

CICTP