| In recent years,with the progressive stabilization in China’s economic development,automobiles have been chosen by the vast majority of Chinese families as a means of transportation,which has brought about a steady increase in automobile ownership in China.But in the meantime,all kinds of traffic problems are emerging thereupon day by day.To resolve the practical traffic problems,vehicle re-identification technology came into being accordingly and entered the public’s field of vision gradually.However,depending merely on single-vehicle attributes such as the painting or vehicle models,etc.,it is scarcely possible to achieve a relatively reliable accuracy in vehicle re-identification.Moreover,the adverse effects from various external factors such as the occlusion of unidentified objects,low recognition,and angle variations in the real scenes,make it excessively arduous to obtain relatively reliable accuracy in vehicle re-identification.Starting from the perspective of solving these disturbances,in the paper,we investigate both the enhancements and optimizations performed on the traditional vehicle re-identification model.The principal work carried out involves as follows:(1)The construction of a vehicle re-identification model dependant on attributes and features based on the lightweight network OSNet.Since the main part of the traditional network employs the first three blocks of the residual network(Resnet-50),not only the basic network embodies excessive parameters,but also branch networks are involved,thus leading to a comparatively intricate structure.Therefore,affected by these factors,the original deep learning model fails to achieve the anticipatory speed in the identifying process of actual training,or any satisfactory training results.On this account,in this paper,a vehicle re-identification framework based on lightweight network OSNet is designed to replace the original Res Net aiming at the disadvantages of the traditional network structure.The method proposed herein was employed in training afterward with the optimized algorithm framework and network model.Judging from the final experimental data,it is evident that,compared with the original model,the vehicle re-identification network model based on OSNet has remarkably elevated the accuracy in training.(2)The construction of a vehicle re-identification algorithm based on multi-task learning.One of the most commonly applied methods for traditional vehicle re-identification is to extract the features of a target with a single convolutional neural network,in which,complex tasks are customarily decomposed into multiple irrelevant tasks for learning respectively afterward.The peculiarity that distinguishes Multi-task learning from a single convolutional neural network is that Multi-task learning enables the gathering of training data from various tasks.In the framework proposed in this paper,each branch is utilized for attribute identification.Judging from the overall structure of the algorithm model,it can be classified as a vehicle re-identification model containing three branches.Subsequently,the framework further collaborates with the model parameters employed in the simultaneous training of multiple different tasks,thus not only improving the learning ability of the model but also expanding the generalization of the model.The training results are then optimized through three successive steps,the approach of complement objective training(COT)from the combination of cross-entropy and complement entropy,the even distribution of the probabilities predicted from other classifications to each category,and then the most possible maximization of the probability of correct predicting on categories.The experimental results show that,compared with the original algorithm that merely utilizes the cross-entropy loss function,the method proposed herein efficaciously improves the identifying accuracy of the vehicle re-identification with the assistance of the data set Ve Ri-776.Meanwhile,the algorithm in the paper has undergone a transverse comparison with other algorithms.Through the final experimental data,it can be seen intuitively that the more outstanding identification capacity is achieved both in improved algorithm framework and enhanced network model,and the method is thus proven efficacious. |
