Research On Vehicle Reidentification Based On Deep Learning

With the use and popularization of automobiles,a large number of technical problems related to vehicle management and scheduling have emerged to be solved.As one of the technical difficulties related to vehicle management and scheduling,vehicle re-identification has been highly concerned by researchers in the industry.In recent years,the vehicle re-identification algorithm based on deep learning has been widely used in the field of vehicle re-identification due to its unique adaptability and strong recognition accuracy.This paper aims to solve the problem of large intra-class difference and small inter-class difference in vehicle re-identification,and expects to improve the generalization ability and network accuracy of vehicle re-identification network.Based on the above problems,this paper uses convolutional neural network as a feature extractor to replace the traditional vehicle re-identification method,and proposes two vehicle reidentification networks based on deep learning,namely,vehicle re-identification network based on multi-granularity and nonlocal relationship and vehicle re-identification network based on dimension decoupling and nonlocal relationship.The m AP of the two vehicle re-identification solutions proposed in this paper reached 78.9 % and 81.8 % on the mainstream dataset Ve Ri-776,respectively.Both solutions met the standards of the mainstream solutions in this field at the present stage,which proved the effectiveness of the proposed solutions.The following is the specific elaboration of the research results of this paper :(1)The vehicle re-identification network based on multi-granularity and non-local relationship proposed in this paper solves the following two problems.First,the training of images will lead to the over-fitting problem when the network extracts useless information.Second,when there are large intra-class differences and small inter-class differences,the network accuracy will be reduced.The network is divided into non-local relationship feature extraction module and multi-granularity feature extraction module,which are used to solve the above problems respectively.(2)In view of the large amount of calculation in the non-local relationship feature extraction module,the interference problem caused by useless non-local relationship on feature extraction,and the bottleneck problem of network accuracy growth caused by hardening sub-strategy defects,this paper innovatively proposes a vehicle recognition network based on dimension decoupling and non-local relationship.The network is divided into global feature extraction module,non-local relationship capture module and dimension deconstruction module.The global feature extraction module is used to extract global features,and the latter two are used to solve the above problems,respectively,to improve network accuracy and network generalization ability.