Research On Vehicle Re-Identification Method Based On Multi-Dimensional Attention Mechanism

In recent years,vehicle re-identification has always been a research hotspot in the field of computer vision,and is an indispensable component of the construction of intelligent cities.The vehicle re-identification task aims to retrieve the same vehicle from multiple cameras.The challenge faced by this task is that factors such as perspective,lighting,occlusion,and resolution may cause significant differences in the appearance of vehicles with the same identity;The appearance difference of vehicles with different identities of the same vehicle type is small.The key to solving this challenge is to capture the vehicle’s identifying information.Attention mechanisms have been widely used in the field of computer vision,such as image classification,target detection,semantic segmentation,and so on,because they enable network models to focus on the main information of the target rather than secondary information.Therefore,the research work in this paper mainly focuses on using attention mechanisms to improve the performance of vehicle re-identification.(1)Aiming at the problem that it is difficult for attention mechanism to finely measure the importance of vehicle semantic information from channel or spatial dimensions alone,this paper proposes a dual-relational attention model for vehicle re-identification.First,a dual-relational attention module is proposed.The module constructs the importance of a feature point in both spatial and channel dimensions through the relationship between feature points,and then models the attention of feature points in three-dimensional space to improve the performance of the attention mechanism and to mine discriminative semantic information more accurately.Secondly,a three-branch network is designed based on the dual-relational attention module to extract the discriminative information of vehicles.In addition,a non-similarity constraint is designed to allow the two branches with the attention module to learn diverse and complementary discriminative information.(2)Aiming at the problem that how to use multi-axis interaction of information to promote effective learning of attention mechanisms,and how to capture important details in local areas of vehicle images,this paper proposes a multi-dimensional attention model for vehicle re-identification using multi-axis interaction.Firstly,a Window-Channel Attention Module is designed.The module configures unique channel attention for each region to extract location level features;Within each region,the learning of channel attention is facilitated through information interaction across locations and then channels.Secondly,a Channel Group-Spatial Attention Module is designed.This module configures unique spatial attention for each channel group to emphasize the location of each part in the spatial dimension;Within each channel group,spatial attention learning is promoted through cross channel interaction followed by cross location interaction of information.Finally,a multi branch network is designed based on these two modules.Two branches of the network extract semantic features at the site level through window partitioning in a priori manner and through channel semantic aggregation in an adaptive manner.These two methods complement each other,improving the network’s feature representation capabilities.Extensive experimental results show that the two models proposed in this paper for vehicle re-identification achieve promising mAP,Rank-1 and Rank-5 accuracies on the VeRi-776 dataset and Rank-1 and Rank-5 accuracies on the VehicleID dataset.