Vehicle Re-identification Research Based On Dynamic Alignment

In recent years,with the development of urbanization,the number of vehicles has been increasing tremendously,and vehicle management has become more and more challenging.At the same time,with the rise of smart cities,the Internet of Things,big data,and 5G,there are more and more connected monitoring devices,resulting in massive data.This also facilitates vehicle monitoring,vehicle tracking and vehicle identification.Therefore,the tracking of the inspected vehicles,the identification of specific vehicles,the monitoring of community vehicles,etc.have become hot research topics,and the related research on vehicle re-identification has also received more and more attention.Vehicle re-identification(Vehicle ReID)is an image retrieval technology.The purpose is to use computer vision technology to identify specific vehicles from the images or video sequences of different cameras.At present,the license plate is the most reliable information for vehicle identification,and the license plate detection is also the most used method in vehicle identification.However,license plate information is not always valid in reality.Due to the effects of lighting,camera angle,occlusion,and dirt,the license plate information sometimes cannot be effectively extracted,which af fects the accuracy of vehicle identification.Vehicle re-identification uses computer vision technology and deep learning to extract and analyze the global and local features of the vehicle,so as to initially identify a specific vehicle.Then,combined with methods such as license plate recognition and manual recognition,the overall effect of vehicle recognition can be effectively improved.This is exactly the application value of vehicle re-identificationThe thesis applies Dynamically Matching Local Information(DMLI)method to local features learning,which can dynamically align horizontal stripes without additional supervision.DMLI can improve the learning effect of local features to a certain extent.The thesis uses a re-identification framework named AlignedReID++,which learns a global feature associated with a local branch based on DMLI.In the local branch,the local features are aligned by introducing a shortest path distance.This local branch can guide the global branch to learn more about the discriminative global features.In the inference stage,combining global feature and local features using DMLI can further improve the accuracy.AlignedReID++uses the Triplet loss with Hard Example Mining(TriHard loss)as the metric loss,and combines Softmax loss and TriHard loss to accelerate convergence.In addition,the better global feature extraction ability makes the method attractive for the deployment of large ReID systems,without the need for expensive local feature matching.In the thesis,the main parameters of the algorithm model,such as pre-loaded network,optimization algorithm,initial learning rate,and learning rate scheduler are fine-tuned.At the same time,the training techniques of reranking and label smoothing are combined.The thesis adopts a new ensemble learning method of Stochastic Weight Averaging(SWA).This method uses relatively constant learning rates to execute the optimizer.After each training cycle,the explored weight combinations are averaged to ensure that as many points(weight combinations)as possible can be explored in the weight space.This will help the algorithm model to jump out of local minima,make the objective function find a better solution in the loss plane.and further improve the generalization ability and accuracy of the model.It is important that this method has almost no additional computational cost and is relatively simple to implement.In order to solve the problem of lack of datasets for vehicle re-identification task in community scenarios,a large-scale related dataset has been collected and produced in the thesis,named Oeasy-Parking.Experiments are performed on the two open source datasets VeRi-776 and VehicleID,and reference experiments are performed on the Oeasy-Parking dataset.The results of benchmark model experiments and improved algorithm experiments are compared and analyzed in detail.At the same time,the thesis also compares with the latest research results to verify the effectiveness of the improved algorithm.The comparison results show that the mAP and rank-1 accuracy of the improved algorithm in the VeRi-776,VehiclelD datasets are higher than the current latest results,respectively.Moreover,97.3%mAP and 99.2%rank-1 accuracy were obtained on Oeasy-Parking dataset.The experimental results have proved the effectiveness of the improved algorithm.The research in the thesis also has insufficiencies such as large algorithm models,lower training speed,and limited number of samples in the datasets.Therefore,there is still a gap from practical applications.Further research on these issues is encouraged.