Research On Target Detection Algorithm For Traffic Night Vision Scene

As one of the important research contents of computer vision,target detection is widely used in industry,military,intelligent monitoring,face recognition,multi-target tracking,automatic driving technology and other fields.At present,with the development of artificial intelligence and autopilot technology,object detection based on deep learning has become a research hotspot of scholars at home and abroad.The network model of YOLOv3 based on Darknet53 feature extractor has the advantages of high detection accuracy and fast detection speed,but there are still some problems,such as weak detection ability for small targets,inaccurate boundary box positioning and large model weight file.Therefore,a target detection algorithm based on improved YOLOv3 model is proposed.The feature extraction module of YOLOv3 is changed from Darknet53 to Vo VNet,which has the advantages of reducing the amount of computation and speeding up the forward propagation speed.The loss function of bounding box in YOLOv3 is changed from Io Uloss to GIo Uloss,which is helpful to improve the ability of bounding box localization.In addition,in order to enhance the classification and localization ability of the network model,the anchor parameters corresponding to the data sets are adapted when training with different data sets.Experiments on open datasets VOC2007 + VOC2012 and Visdrone2018 show that the proposed algorithm improves the map of target detection without affecting the real-time performance of YOLOv3 algorithm.The map trained on open datasets Visdrone2018 is 40.9%,which is 8.7% higher than that of traditional YOLOv3;The map trained on VOC2007 + VOC2012 is 81.6%,which is 3.5% higher than that of traditional YOLOv3;After training and testing on the traffic night vision scene data set made by the author,the map can reach 84.1%.Meanwhile,the weight file is reduced from 234 M to 163 M.In addition,the influence of different loss functions on the performance of the network model is explored.At the same time,in order to further reduce the weight file size of the network model,a lightweight target detection network model based on improved YOLOv3 is proposed by changing the feature extractor of YOLOv3 to the Ghost Net module and changing the bounding box loss function from Io Uloss to GIo Uloss.Experiments on the open data set Visdrone2018 show that the map of the improved lightweight model is 14.28% higher than that of YOLOv3-tiny;After training and testing on the traffic night vision scene data set made by the author,the map can reach 95.3%.Meanwhile,the weight file is reduced from 234 M to 89.9M.In addition,the effects of different loss functions and input image scales on the performance of the network model are explored.