Hierarchical Vehicle Identification System In Monitoring Scene Based On Deep Convolutional Feature Fusion

Vehicle automatic identification technology is an important research topic in intelligent transport system.Its purpose is to solve the problem of identification of specific vehicles in the actual traffic monitoring scenarios and retrieval of target vehicles in massive databases for improving the efficiency of traffic management and public security law enforcement.There are rich visual appearance attributes of vehicles and this paper mainly utilize the license plate character,license plate color,vehicle logo,pasted marks,vehicle model and vehicle color to study the identification of specific vehicles in traffic monitoring scenes,including license plate detection and recognition and vehicle re-identification.The main research work is as follows:(1)An improved license plate detection method based on Fast R-CNN through multi-characteristic and context fusion is proposed.Firstly,to improve the license plate detection performance,we introduce a region called context-of-plate combined with vehicle location as the context information,exploiting the hidden correlation.And to extract local and contextual features,we make some modifications to the Faster-RCNN: select several layers with different shades of VGG16 for vehicle,context-of-plate and license plate to obtain multi-scale integrated feature maps,complementing location and semantic information,enhancing the representation of features and reducing scale interference.Then,the features of license plate region and the context region are fused to refine the license plate detection.Experiment results on benchmark datasets demonstrate that the accuracy of the proposed method in detection of various license plates exceed 98%.(2)A vehicle re-identification algorithm that fuses global and local features is proposed.It utilizes the pre-trained targeted neural networks to extract the corresponding vehicle features separately.Firstly,the global features including vehicle color and model and the local features including vehicle logo,light and window are extracted respectively.And then the feature relationship between global features and local features is mined through feature mapping and feature fusion under the guidance of part attention model.Finally the Siamese neural network is used to share the weight parameters for the two sets of fusion features,calculating their distance in Euclidean space,which is used to judge the similarity degree of two input vehicle images.Experiments on Vehicle ID and Ve Ri datasets indicate that the proposed algorithm shows its superiority in both CMC and m AP indicators.(3)Adaptive network for different features and spatial-temporal re-ranking are introduced into vehicle recognition algorithm.Considering that VGGNet is used as feature extraction network for all features,some feature information extraction may be incomplete,while some feature extraction will generate unnecessary calculation due to the deep layer,different adaptive networks are adopted for different features.For the vehicle color features,Alex Net is used,the model features are sensitive to Goog Le Net,and all local features are extracted by Dense Net.With the help of the spatial-temporal information of the vehicle image as priori knowledge,the optimal spatial-temporal path is generated through the chain MRF model.And the feature matching is carried out in the Siamese network to determine the similarity between the input couples,optimizing the vehicle re-identification.The experimental results show that the algorithm with adaptive network and spatial-temporal re-ranking reduces the computational complexity of the network as well as improves the accuracy of vehicle re-identification.