Vehicle Re-Identification Based On Multi-Scale Attention Mechanism And Information Fusion

Vehicle Re-Identification(Re ID)technology can achieve cross-view retrieval of target vehicles,and this technology has important application prospects in the fields of intelligent transportation,intelligent monitoring and multi-objective video tracking.In recent years,with the development of deep learning,vehicle re-identification methods based on deep learning have gradually become an important research direction.However,due to the different data collection devices and the diversity of actual traffic scenes,the collected vehicle images have problems such as large intra-class variability,high interclass similarity and different scales,and the deep learning-based vehicle re-identification methods are still insufficient for such problems.This thesis focuses on the research of vehicle re-identification method based on vehicle appearance information,focusing on the salient feature extraction for vehicle appearance and the vehicle re-identification method that fuses multi-scale features.The main research contents are as follows:(1)A vehicle re-identification method based on multi-scale attention mechanism was proposed.The method aims to solve the problem of large intra-class variation and high inter-class similarity of vehicle images.Based on the salient feature extraction performance of the attention mechanism,this thesis firstly,constructed a multi-scale spatial attention module to realize the spatial feature weight assignment to different scales,and focuses on the channel features highly relevant to the vehicle re-identification task based on the channel attention,so as to realize the spatial and channel salient feature weighting;then,jointly uses the triadic loss and cross-entropy loss functions to guide the training of the model;finally,the re-ranking algorithm was introduced to reorder the initial sequences predicted by the model,and the re-ranking sequence results can effectively improve the prediction performance of the re-identification model.(2)A vehicle re-identification method based on multi-scale information fusion was proposed.The method aims to solve the problems of different scales of vehicle images and different proportions of vehicle targets in the figure,and considers the importance of different scale information in both the multi-scale feature extraction and feature fusion stages.Firstly,a multi-scale feature extraction module was proposed in the feature extraction stage,which uses the features under different stages in Res Net-50 as a multiscale feature pyramid,extracts the features at different scales and merges the semantics of high and low-level features,and sends them to the multi-scale spatial attention module to realize significant feature weighting.Secondly,a complementary multi-scale feature fusion module was proposed in the feature fusion stage,which jointly uses dilated convolution with different dilation rates and deformable convolution with different size convolution kernels to achieve more robust features by fusing multi-scale feature information of wide field of view,which is used to improve the performance of vehicle re-identification models.The method considers the importance of different scale information in both the feature extraction and feature fusion stages,which can effectively solve the problems of different scales of vehicle images and different proportions of vehicle targets in the figure.The proposed method was validated by ablation experiments and comparison experiments on the publicly available vehicle re-identification datasets Vehicle ID and Ve RI-776,respectively,and the results show that the proposed method can effectively improve the prediction performance of the vehicle re-identification model.