Research On Vehicle Re-identification Based On Multi-Scale Feature Fusion And Cross-Domain Adaptation

With the continuous increase in the number of motor vehicles,the need for computer technology to assist manual traffic management has become very urgent in recent years.Vehicle Re-Identification(Re-ID)has become an important means to locate,track,and supervise vehicles in the city,and has been widely researched by the academic community.Benefiting from deep mining of deep image features,the vehicle Re-ID technology has been well improved in performance and accuracy.Nevertheless,there are still some problems that hindering the further development in this area.The first problem is that in the case of sufficient labeled data,the ideal results can often be achieved through supervised manner.However,the vehicle images with low illumination fail to extract representative vehicle features as well as the images with normal illumination.Besides,the complexity of network models makes it difficult to transfer training models to mobile terminals and embedded devices.The second problem is that the existing unsupervised methods for vehicle Re-ID mainly use cross-domain adaptation to alleviate the shortage of labeled data,however it is difficult to solve the domain gap between different datasets and pseudo label noise.To settle the poor model performance in low illumination scene and difficulties in porting to mobile and embedded devices in question one,this paper proposed a vehicle re-identification method based on lightweight multi-scale feature fusion network in low illumination scenes.The method includes Image Enhancement method with Dual Frequency Domain(IEDFD)module and Lightweight Multi-scale Feature Fusion Network(LMFFN)module.Firstly,IEDFD enhances the image quality by amplifying the high-frequency component and improving the illumination of the low-frequency component to optimize the accuracy of model feature extraction.Secondly,LMFFN obtains multi-scale features through multi-branch networks and reduces the complexity and parameters of the model by using the small-scale convolution kernel and feature reduction operation.To resolve the domain gap and pseudo label noise in unsupervised cross-domain adaptation in problem two,this paper proposes an unsupervised vehicle Re-ID method based on cross-domain adaptation and feature cross-division.The method consists of two stages.In the first stage,Cross-style Semi-supervised Pre-training(CSP)is designed.CSP jointly trains Re-ID model through the source domain and cross-domain style samples by a semi-supervised manner.Simultaneously,a pseudo label reassignment strategy is proposed to allocate reasonable soft labels to generated data.In the second stage,the fine-tuning based on Feature Cross-Division(FCD)is designed to reduce the misleading of pseudo label noise by increasing the confidence of pseudo label.Based on the PyTorch deep learning framework and the public vehicle datasets,this paper has carried out a lot of tests and experiments on the above methods.The results show that the first method proposed in our paper can effectively improves the accuracy of the vehicle Re-ID method in low illumination scenes.First,this method improves the illumination of the images and obtains multi-scale features through lightweight multi-scale feature fusion network,which increases the diversity of features.Second,this method reduces the amount of network parameters and complexity of the model by designing lightweight multi-branch network under the condition of ensuring performance.Meanwhile,the second method can alleviate the restriction of domain gap on the generalization ability of the pre-training model,and effectively suppress the impact of pseudo label noise on the accuracy of Re-ID by boosting the confidence of the pseudo labels.Compared with the existing methods,the above two works have achieved better performance.