Study On Logistic Vehicle Feature Recognition Based On Improved Faster R-CNN

The traditional License Plate Recognition System can not meet the needs of detecting logistics vehicle in the complex logistics industry environment.Logistics is called the 'third-party profit source' of enterprises in logistics industry,especially in logistics transportation etc.Effectively detecting vehicles,obtaining logistics infomation,overall planing and reducing unnecessary costs have became more and more significant and profound meaning.But at present,most vehicle identification systems are designed for cars,which are lacking of research on logistics vehicles.The deep learning algorithm include two modules,the first is object detection network and the second moudels is classification network,Faster R-CNN is selected as basic target location and identification network in this paper.With a large number of pictures on the logistics park scene,we subdivid various types of vehicles data.The data sets are divided into train sets,validation sets,and test sets.For the problem of logistics vehicle model always falling into the local optimal solution,and then the parameter optimization setting research is carried out.(1)The purpose of this research is to achieve accurately detect the characteristics of the target logistics vehicles.As for the problem of training over-fitting,the batch normalization method was camed up.The improved model improve the effect of the training and solve the problem that the model falls into the local optimal solution problem.(2)Aiming at the problem of unbalanced training sample and the special logistics vehicles difficult training,under the condition of retaining the orignal samples,a method based on online hard example mining and hard negative example mining technology is proposed.The maximum loss value of real-time screening in training for error transmission,which solves the problem of sample insuffiency,imbalance and difficult identification of specific vehicle targets,achieves accurate classification of difficult vehicle targets.(3)In order to improve the recall rate and generalization of the model,in the multi-target location detection,as for the problem of the traditional non-maximum suppression missing target,the idea of the confidence penalty function is adopted,and the new non-maximum suppression is improved.The algorithm improves the recall rate of vehicle detection and solves the problem of vehicle target missed detection.(4)Finally,through the sensitivity analysis of the non-maximum suppression,the Gaussian confidence penalty function method is used to improve the recall rate greatly and improve the precision of the model,which lays a firm foundation for the accurate correlation of logistics vehicle information.With the above improvement method,the comparison between before and after retrofit are made.The recall rate of model is increased from 83.5% to 98.5%,the recognition accuracy is increased from 90.6% to 97.8%,and the mean Average Precision(mAP)is increased from 81.2% to 90.3%,which can be presented to prove the superiority of proposed algorithm.The method of this paper is applicable to the training optimization problem of multi-target precise detection and recognition,as well as the correlation of logistics vehicle information,statistics and scheduling of vehicle information.