Vehicle Re-identification Algorithm Based On Deep Learning

Vehicle Re-identification is a technology for identifying target vehicles through images or video sequences,and is also considered a sub-problem of image retrieval.This technology is widely used in intelligent parking,suspect vehicle tracking and intelligent traffic management,etc.,and is of great significance for the construction of intelligent traffic system.Since the license plate is the only identification of the vehicle identity,the license plate recognition technology has been widely used in the vehicle re-identification technology,but in the real traffic environment,there are cases such as forged license plate,blocked license plate and no license plate.Therefore,this paper mainly studies the vehicle re-identification technology based on vehicle attributes and appearance.In practical applications,the same vehicle will be difficult to be retrieved correctly due to factors such as changes in viewing angles,different lighting and resolutions,and different vehicles may be incorrectly retrieved due to the mass production of vehicles.In response to the above problems,this paper conducts in-depth research on vehicle re-identification algorithms,and proposes two vehicle re-identification methods based on different network architectures.The specific work of the thesis is as follows:1)In order to solve the problem of subtle differences between vehicles and changes in viewing angles,this paper proposes a vehicle re-identification method based on local features and viewpoint aware.The method is a dual-branch network structure,in which the global branch extracts global features and uses ID loss and cross-entropy loss for training.In the local branch,a semantic segmentation network and a viewpoint-aware network trained with soft labels are used to construct a local feature module and a viewpoint-aware module.Among them,the local features can extract more discriminative vehicle features,combined with the continuous viewpoint prediction value to give different weights to different local regions,so as to reduce the impact of viewpoint changes on vehicle reidentification,and at the same time solve the problem of large intra-class differences.problems with small differences.Experimental results show that this method can effectively improve the accuracy of vehicle re-identification.In the verification stage,this paper performs a secondary optimization sort on the initial sorting results to further improve the re-identification accuracy.2)Aiming at the loss of detail information caused by convolution and downsampling operations in convolutional neural networks,this paper proposes a vehicle re-identification method based on visual Transformer,using pure Transformer architecture to extract finegrained features of vehicle images.Add an auxiliary information module to solve the problem of viewing angle changes by using the angle of view and camera ID tag information.Aiming at the problem that the Transformer network will cause network degradation and feature similarity increase due to the deepening of the number of layers,this paper proposes a cosine similarity loss and a contrast loss.Through the optimization of the loss function,the network can learn more diverse features.Finally,it is proved by experiments that this method can improve the accuracy of vehicle re-identification.