Traffic light recognition based on one-dimensional convolutional neural network

Changsuk Oh, Dongseok Sim, H. Jin Kim

Abstract: Recognizing traffic lights is an essential task for drivers and autonomous vehicles. In this paper, we introduce the barcode generation method that reduces the dimension of traffic light images, which reduces the computation and memory load for a signal classifier. Through the barcode generation, the traffic light images are reduced to one-dimensional images and classified using a one-dimensional convolutional neural network. Since the proposed module is trained and verified with traffic light images including background, it is suitable to be used in conjunction with various detection algorithms whose output images include background. Compared with two-dimensional recognition modules, the number of parameters is reduced by half, without significant degradation in classification performance.