Research On Night Traffic Signal Recognition Based On Deep Learning

Traffic signal recognition is an important component in the fields of traffic management,autonomous driving and assisted driving.Obtaining accurate and real-time traffic signal information can reduce the occurrence of traffic accidents and improve the efficiency of traffic operation.However,the current traffic signal recognition mainly focuses on daytime,and there is little research on nighttime with more complex backgrounds and more interfering light sources.In addition,existing models have low accuracy and slow speed for nighttime traffic signal recognition.To address the above issues,the paper proposes a nighttime traffic signal recognition model based on deep learning,and the main research contents are as follows:Firstly,to address the issues of few data samples and single weather condition for nighttime traffic signal,the paper proposes an expansion scheme for the nighttime traffic signal dataset based on Cycle GAN(Cycle-Consistent Generative Adversarial Networks).Three sets of experiments were designed,which were sunny to rainy,sunny to foggy,and sunny to snowy.The experimental results show that,the nighttime traffic signal images under different weather conditions simulated in the paper based on Cycle GAN have clear image quality and complete content.And the constructed Night TS-3Plus(Night Traffic Signal-3Plus)dataset effectively solves the problem of imbalanced nighttime traffic signal samples in different weather conditions,providing sufficient data support for the subsequent experiments.Secondly,to address the problems of high noise and high degree of background interference in nighttime traffic signal images,a scheme to add the attention mechanism is proposed with YOLOv5s＿5.0(You Only Look Once version5s＿5.0)as the benchmark model,which in turn enables the model to improve the perception of key features of nighttime traffic signals and minimized the interference caused by noise,so that the model can improve the attention to the nighttime traffic signals and the accurate recognition can be achieved even when the nighttime traffic signals are obscured.Thirdly,to address the problem that the nighttime traffic signals themselves do not have high resolution and thus cannot provide detailed feature information,two improvement schemes are proposed with YOLOv5s＿5.0 as the benchmark model: first,to increase the resolution of the feature map,and to enhance the sensitivity of the model for nighttime traffic signals by removing large-sized detection head and adding micro-sized detection head.Second,to optimize the feature fusion structure.Repeated downsampling will lead to the loss of key feature information.By optimizing the Bi FPN feature fusion network to effectively fuse feature maps from different scales,the perception of nighttime traffic signals can be improved,and then more effective information can be obtained to improve the accuracy of nighttime traffic signal recognition.Finally,to verify the effectiveness of the YOLOv5s＿NightTS model proposed in the paper,an experimental platform is built and the corresponding software and hardware environments as well as evaluation metrics are selected.On this basis,the horizontal comparative experiments of different models and the vertical comparative experiments of different weather conditions were designed.In addition,visualize and record the experimental data for subsequent analysis and research.The experimental results show that,the m AP(mean Average Precision)of the YOLOv5s＿Night TS model proposed in the paper reaches 96.01%,the F1 score is 0.92,the recognition speed reaches 36 f/s,the loss function decreases steadily and tends to converge,which basically meets the requirements of accuracy and real-time of nighttime traffic signal recognition.