Research On Vehicle Detection Algorithm Based On Improved EfficientDet

Vehicle detection using deep learning technology has always been the focus and difficulty of establishing intelligent traffic management system.The algorithm of Vehicle detection in traffic scene be required that it must ensure a good detection accuracy and also achieve good detection efficiency in order to meet the two-dimensional requirements of accuracy and real-time.Therefore,the construction of vehicle detection that can meet the needs of real scenes is of great research significance for the realization of intelligent traffic control.In this paper,after in-depth research on the EfficientDet object detection algorithm,an improved EfficientDet vehicle detection method is proposed,and experiments are carried out on the KITTI dataset after screening and data augmentation to verify the accuracy and efficiency of the new model.In the research of data augmentation methods,the influence of basic data augmentation and mixed sample data augmentation on KITTI dataset is verified on the EfficientDet.Through experiments,the accuracy improvement of six methods:geometric transformations,color space augmentations,noise injection,Mixup,CutMix and Mosaic is analyzed,finally,Mosaic method with the best effect is selected as the data augmentation method,and the dataset using Mosaic was used as the dataset for subsequent experiments to train the model afterwards.In the research of object detection algorithm,aiming at the problem that EfficientDet detection method can not adaptively match the dataset to generate anchors,combined with K-means unsupervised clustering method,a boundary box clustering optimization prediction aspect ratio method is proposed to replace the manual setting method of the original model,so as to further improve the accuracy of the model.Aiming at the SGD with Momentum optimizer used in EfficientDet model,use the AdamW method to execute weight decay after the gradient term transformation operation,and then perform gradient decay when the parameters are updated,so as to ensure that the penalty term value is also large when there is a large gradient and improve the convergence efficiency in model training.To address the current situation that there are many small object in the sample,an improved feature fusion layer structure is proposed,using the method of reducing the number of under-sampling in order to achieve the purpose of extracting detailed feature information,based on this,a cross-level connection is used to fuse the feature of the upper and lower layers,preserving the semantic information of the shallow layers and the details of the objects,making the structure of feature fusion with stronger extraction capability.To address the problem of uneven positive and negative samples and sample categories in the filtered KITTI dataset images,the loss function is optimised and EFL is used to dynamically adjust the loss contribution according to the ratio between positive and negative samples and the category of rare samples to overcome the adverse effect of this problem.On the KITTI dataset using by Mosaic,it is found that the accuracy and reasoning speed of the improved EfficientDet vehicle detection model significantly improved compared with the original model,accomplished the accurate vehicle object detection and meets the requirements of vehicle real-time detection,meanwhile,the improved EfficientDet has better detection performance for small objects,verifing the effectiveness of the data augmentation method and optimization method for model used in this paper.Figure[43]table[9]reference[68]...