Road Traffic Target Detection And Congestion Prediction Based On Deep Learning

With the rapid growth of domestic car ownership,the problem of traffic congestion is becoming more and more serious,which brings a lot of inconvenience to people’s production and life,and also threatens the safety of people’s life and property.How to effectively identify and predict the traffic state urgently needs a solution.Computer vision technology based on deep learning can effectively construct a complete set of traffic congestion status discrimination methods,which has high research value.In this paper,based on the deep learning method,the target detection of traffic participants is carried out,and the traffic parameters such as traffic volume are accurately extracted by the target tracking method,and the traffic congestion status is further discriminated and analyzed by the short-term traffic prediction method.The paper first built a deep learning data set based on traffic scenarios,and analyzed the deep-level features of the data sets in traffic scenarios.Based on this,some improvement ideas were proposed,and a new deep learning network model was constructed.Through comparative experiments,YOLOV5 is selected as the target detection algorithm in this paper.There are four specific improvements: The first is to improve the adaptive anchor frame clustering method,select a new clustering method,and add a genetic algorithm to improve the clustering accuracy and the matching accuracy of the initial anchor frame;The second is to improve the number of anchors in the adaptive anchor frame,reduces the number of anchors,improves the inference speed without affecting the detection accuracy,and accelerates speed of model convergence;The third is to optimize the Mosaic data enhancement.According to the characteristics of the self-built data set,the method of Mosaic data enhancement is redefined,and a new Mosaic data enhancement is constructed,which further improves the convergence speed of the algorithm;The fourth is to improve the multi-scale prediction of FPN.A new Half-FPN multi-scale prediction structure is proposed for image features.By reducing the amount of calculation and parameters,the inference speed is improved,the detection speed is accelerated,and the detection accuracy is improved.The clustering accuracy of the final model algorithm was improved from 71.6% to 85.1%;the model convergence speed was shortened from 54.12 h to 15.45h;the detection accuracy was improved from 74% to 81.2%;the detection speed was shortened from 13.9ms to 12.9ms.This paper builds a traffic congestion discrimination datasets.The LSTM model is selected as the basic algorithm for prediction,and the algorithm model is improved and optimized.The specific improvements are in the following two aspects: The first is to improve the loss function of the LSTM model,adding a classification loss function to the prediction task,which improves the learning ability of the model and improves the accuracy of prediction;The second is to optimize the structure of the LSTM forgetting gate and add a classification item,which further strengthens the model’s ability to process boundaries when judging and improves the accuracy of prediction.This paper realizes the discrimination and multi-step prediction of short-term traffic state,and improves the prediction accuracy.The prediction accuracy has increased from the original 91.67% to95.56%In this paper,a clear threshold standard for traffic state discrimination is given.After the detection results are obtained through the improved YOLOV5 algorithm,the basic information of the traffic flow is obtained through the multi-target tracking model Deep Sort.Based on the obtained traffic flow information,an improved LSTM prediction model is established to predict the traffic congestion state.So far,a complete solution has been constructed,and the discrimination accuracy and prediction accuracy of the model have been improved.The complete traffic state prediction scheme finally obtained in this paper can provide useful reference and reference for relevant traffic management departments.