| Vehicle Re-identificaiton is a new emerging technology in the field of intelligent video analytics,which aims to accurately match the vehicle identities under nonoverlapping cameras in complex video environment and determine whether the vehicle images captured by different cameras at different moments are the same as the given target vehicle,thus the manpower and time needed for accurate retrieval of vehicles can be effectively saved,and the detection efficiency can be improved.However,vehicle license plate information is difficult to obtain from video due to such factors as shooting orientation,low resolution,shielding and uncontrollable environment in video of urban road network monitoring of real scene.So vehicle image feature matching is required to locate the target vehicle quickly.This thesis combines feature learning and metric learning,designs a multi-task learning based vehicle feature representation model,and builds a vehicle re-identificaiton system with high accuracy based on the actual road network environment.The content of this thesis is as follows:1.Research on collaborative metric learning strategy based on multi-task.The high robustness image feature representation learned from the classification task combined with the clustered attributes of characteristic subspace which is obtained through feature transformation in metric learning are proposed by analyzing the correlation between the classification task and the metric learning in the convolutional neural network.Then the model learning is carried out for the constraint auxiliary measurement of feature representation by classification,which solves the problem that the metric learning model is difficult to train and the training effect is poor.The experimental result shows that the multi-task learning strategy is more accurate than the single metric learning in the feature representation of vehicle images and improves the accuracy of vehicle recognition.2.Research on significant local feature representation of vehicles.Significant feature enhancement model and adaptive local feature selection mechanism are proposed.Besides,a feature mapping model based on the fusion of global and local features is designed.This model solves the contradiction between the need for global feature representation for vehicles of different vehicle colors and the need for strong discriminant feature representation for vehicles of similar vehicle colors in vehicle re-identificaiton by multi-granularity feature mapping of vehicle images.Experiment on the open data set shows that the model can achieve the highest accuracy of vehicle re-identificaiton.3.A complete vehicle re-identificaiton system is designed and implemented.The system architecture is designed by combining the characteristics of low delay and high accuracy in the practical application of vehicle re-identificaiton.A complete demand analysis,system design and system implementation are carried out,and the HBase is used to build the vehicle index database,so that the system can quickly build the retrieval database in the massive vehicle images and then respond to the vehicle re-identification request.Finally,the efficiency and accuracy of the vehicle re-identificaiton system designed in this thesis are verified by the functional test and performance test of the system.The research results can be used in the intelligent monitoring system,which has positive significance for the smart city,smart security and other related fields. |
