Two-stage Target Detection Algorithm Based On YOLOv4

The arrival of the era of big data has changed the prospects of many fields of science and engineering.The transportation field also uses big data to promote the popularization of intelligent transportation dynamic monitoring systems.Road target detection is an important part of intelligent transportation dynamic monitoring systems,which is of great significance for promoting the progress of intelligent transportation dynamic detection systems.However,the complex and changeable road dynamic detection scenarios have brought enormous challenges to the target detection technology.The current mainstream target detection algorithms have problems such as missed and erroneous target detection,and the imbalance between detection accuracy and speed performance,making it hard to content the demands of intelligent traffic dynamic monitoring systems for road target detection algorithms.Therefore,based on the deep learning target detection technology,this dissertation conducts research on road target detection algorithms for dynamic monitoring of complex traffic scenes based on the YOLOv4 algorithm,and proposes a YOLOv4-VT two-stage target detection algorithm.The specific research content includes two points:(1)An improved YOLOv4 model is proposed to improve the detection accuracy of target detection models while bringing higher model complexity.To solve the problem of feature information loss caused by unreasonable weight allocation of the channel feature graph of the YOLOv4 algorithm,the channel attention mechanism ECA is added to the feature extraction network,making the model more sensitive and critical for extracting information within the channel;The spatial pyramid pooling layer of the original model is compared with different pooling scales to select the most suitable pooling scale for this model;To solve the coupling problem of the YOLOv4 detection head,a decoupling detection head is used to replace the original coupling detection head;Aiming at the problem that the CIOU loss function used in the YOLOv4 algorithm cannot handle the serious imbalance between positive and negative samples,a Focal-EIOU loss function is proposed to replace it to improve the detection accuracy.Furthermore,a new YOLOv4-ECA model is constructed to improve the detection accuracy without increasing the model size.(2)Aiming at the different degrees of problems such as target classification error detection and missed detection in the YOLOv4-ECA model in target detection,which makes it difficult to combine the two tasks of target location and target classification,a two-stage target detection algorithm YOLOv4-VT based on the fusion of YOLOv4-ECA target location network and Vision Transformer target classification network is proposed.For target location networks,The original classification and prediction related parts in the YOLOv4-ECA framework have been deleted from the network,making the YOLOv4-ECA target positioning network more focused on completing the positioning task of the target on the image.In the classification task section,the Vision Transformer model and the method of setting classification thresholds are used to solve the identification problems such as target misdetection and missed detection that are prone to occur.Based on the analysis of the experimental results,it is based on the highway live data set.It has been verified that the YOLOv4-VT road target detection algorithm proposed in this paper can solve the problems of missed and erroneous detection of targets,as well as the imbalance between detection accuracy performance and detection speed performance,and meet the requirements of intelligent traffic dynamic detection systems for road target detection algorithms.Figure 46 table 18 reference 68...