Research On Traffic Light Detection By Deep Neural Network Based On Saliency Feature Preprocessing

Traffic signal light detection is an important branch of object detection.Recognize traffic signal lights quickly and accurately plays an important role in assisted driving and unmanned driving.It can provide protection for people’s safety and reduce the occurrence of traffic accidents.The image generated by the vehicle-mounted camera is usually greatly affected by the light,while the detection speed of the traditional object detection algorithm is very slow.Therefore,how to reduce background interference and speed up detection has become the focus and difficulty of research.Aiming at the main characteristics of traffic lights,this paper uses HSV color segmentation and traffic signal geometric feature filtering to preprocess the image,reducing the impact of illumination on the image,and filtering part of the background information in the image with morphology.In order to improve the speed of detection and recognition of traffic lights,deep learning method is adopted as the classifier of the model.This paper improves the YOLOv3 algorithm,which is mainly aimed at detecting and recognizing traffic lights.First,use the anchor frame suitable for the size of traffic lights generated by linear scaling-based K-means clustering,and at the same time improve the feature pyramid network,reduce the number of convolutional layers at large and mediumscale outputs,and enhance its feature extraction capabilities for small targets such as traffic lights and simplify the model structure.Improved Mosaic algorithm use to enhance training data to increase the number of samples in categories with less data.Finally,a deep neural network model based on saliency feature preprocessing is designed for the detection and recognition of traffic lights.This model combines the saliency feature preprocessing module designed for traffic lights and an improved deep learning classifier module.Through experimental tests on the self-made CQTLD dataset and Lara dataset,the overall evaluation of the designed model is carried out.Experimental results show that the improved YOLOv3 algorithm designed in this paper has improved accuracy,recall and detection speed compared to YOLOv3 and YOLOv4 algorithms.The final model designed in this paper has an accuracy rate of 99.58% and a recall rate of 98.47%,which not only improves the detection accuracy,but also reduces the number of false detections of the model.And the detection rate can reach 43 FPS,which can meet the requirements of realtime detection of traffic lights.