Research on Traffic Signal Light Recognition Method in Complex Scene

Aiming at the problem of poor real-time performance and low recognition rate of traffic light recognition methods in complex scenes, a traffic light recognition method for complex scenes is studied. Firstly, the bilateral median filter is used to suppress the noise in the complex background, and the image with low illumination is enhanced; Then, the improved Scharr operator is used for edge segmentation, and the interference region is removed by mathematical morphology processing and setting double threshold, and the accurate location of signal lamp is realized by region analysis algorithm; Finally, according to the color histogram of Cr component in YCbCr color space, the threshold range for judging color is set, and according to the shape characteristics of signal lamps, the discrimination conditions of shape or number is set. Finally, the accurate recognition of traffic signal lamps is realized. The experimental comparative analysis shows that for the traffic light recognition of complex scenes such as cloudy, rainy, snowy and foggy days, the average recognition accuracy of this method is 98.8%, and the average running time of the algorithm is 0.8s. The recognition rate is greatly improved, and the traffic light can be recognized quickly and accurately. Therefore, this method has good robustness and real-time performance.

I. Introduction

As one of the key technologies of intelligent transportation and intelligent driving, traffic signal recognition provides a basic guarantee for the practical application of intelligent driving and unmanned driving, and has become a research hot spot in the field of unmanned driving. Pan et al. [1] have realized the detection and recognition of traffic lights based on Faster Region-Convolutional Neural Network(Faster R-CNN), and verified the effectiveness of the method in real traffic scenes. Qian et al. [2] used the deep learning algorithm to realize the rapid detection and recognition of traffic lights, which overcomes the technical difficulties such as the complex and changeable scene of traffic lights detection and the small proportion of targets in the images of the detection data set. By reducing the number of convolution layers in the backbone network, updating the CSP residual structure and reducing the number of detection scales, Deng et al. [3] proposed a traffic signal recognition method based on the improved YOLOv5s model, which improved the recognition accuracy and real-time performance. Ma [4] introduced the attention module into the deep neural network, increased the weight of the region of interest of the feature map, and proposed a two-way regression based on the attention mechanism and a cross view oriented traffic signal detection method, which solved the problems of large amount of network calculation and difficult deployment of training model at the embedded end in practical applications. Li et al. [5] used convolution neural network and gradient context feature to realize signal lamp detection, which reduces the false detection rate. Zhao et al. [6] proposed a traffic light recognition method based on improved YOLOv4, which solved the problems of low detection accuracy and poor real-time performance for small targets, improved the recognition accuracy of traffic light status information, and reduced the model size.