Research On Mixed Traffic Object Detection Algorithm Based On Deep Learning

In recent years,with the development of society,China's car ownership continues to grow,and the annual traffic accidents and casualty rate are also high,which brings huge economic losses.Intelligent transportation system can effectively collect traffic information and transform it into structured information,which brings great convenience to traffic management.The most important part is the accurate detection of the mixture of pedestrian object and vehicle object.The existing road traffic scene video structured system is often restricted by two conditions.Firstly,at present,the open-source target detection algorithm has a good detection effect in various scenarios,but the detection accuracy still can't meet the demand in the actual scene,and also can't meet the real-time detection.Secondly,the detection results of road traffic scene are greatly affected by the weather,which often occurs in the scene with sufficient light,and the detection accuracy is high in the scene with poor light conditions such as dusk or rainy.Based on these two problems,the author analyzes the road traffic scene and traffic target,and puts forward an efficient detection network of deep fusion of feature map for road scene target detection according to its characteristics.In addition,the existing image enhancement algorithm is studied,and the Retinex algorithm and the accuracy of image and model detection are constructed to enhance the robustness of the detection model.The main work and innovations of this paper are as follows:Firstly,aiming at the problem that the open source data set does not meet the actual needs,the road traffic object data set for the monitoring scene is constructed.This paper obtains the video data of multiple road monitoring and marks them manually.The traditional method of data augmentation combined with the method of Generative Adversarial Networks is used to enrich the data,and the pedestrian and vehicle data set for road monitoring scene is constructed.Secondly,in view of the low accuracy of the target detection algorithm due to the lack of light,this paper proposes to effectively link the information of the scene to be detected with the Retinex image enhancement algorithm and build the enhancement model.Experiments show that Retinex has different enhancement effects on different lighting scenes.In order to maximize the enhancement effect of the image,this paper estimates the average gray value of the scene and adjusts the setting of key parameters.Through repeated experiments,the image enhancement algorithm and the scene information are finally modeled,and good enhancement effect is achieved.Finally,aiming at the problem of uneven size distribution of the changeable targets in the background of road traffic scene,the detection network DF-Net based on the deep fusion of the feature map is proposed through the statistical analysis of the monitoring scene and the detection object,as well as the design of the convoluted feature map visualization results.DF-Net transfers the size of the feature map at the top of the feature pyramid and merges it with the feature map at the bottom.It merges the deep and shallow information,improves the ability of feature extraction,and improves the accuracy of model detection combined with multi-scale prediction.The results show that the accuracy and speed of DF-Net are significantly improved.In addition,the scene of viaduct,pedestrian and vehicle intersection sidewalk and congested highway with small target pixel are measured.The experiment shows that the detection effect and speed of DF net meet the actual needs.