Real-time method for traffic sign detection and recognition based on YOLOv3-tiny with multiscale feature extraction

Abstract

As a part of Intelligent Transportation System (ITS), the vehicle traffic sign detection and recognition system have been paid more attention by Intelligent transportation researchers, the traffic sign detection and recognition algorithm based on convolution neural network has great advantages in expansibility and robustness, but it still has great optimization space inaccuracy, computation and storage space. In this paper, we design a multiscale feature fusion algorithm for traffic sign detection and recognition. In order to improve the accuracy of the network, the gaussian distribution characteristics are used in the loss function. The training and analysis of two neural networks with different feature scales and YOLOv3-tiny were carried out on the Tsinghua-Tencent open traffic sign dataset. The experimental results show that the detection and recognition of the targets by networks with multiple feature scales have improved significantly, and the recall and accuracy are 95.32% and 93.13% respectively. Finally, the algorithm of traffic sign detection and recognition is verified on the NVIDIA Jetson Tx2 platform and delivers 28 fps outstanding performances.