Research And System Implementation Of Vehicle Re-identification Algorithm Based On Deep Learning

Recently,the number of vehicles has increased,making manual-based traffic control increasingly difficult.There is an urgent need to introduce intelligent management tools,hence the emergence of the intelligent transportation system.As one of the important technologies of intelligent transportation hence the emergence of the intelligent transportation system,vehicle re-identification technology can identify specific vehicles in images or videos with different perspectives,playing an important role in intelligent security and vehicle tracking.At present,relying on license plate recognition is the most reliable re-identification method,but due to the existence of variable environments and different shooting locations leading to easy obscuration of license plates,fake license plates occur,number plate information involves the owner’s privacy protection and other issues,so the research of vehicle re-identification technology for vehicles without license plates is very important.The purpose of this paper is to study the use of deep learning technology to mine vehicle surface features for vehicle re-identification.And there are two problems in the existing methods: 1.Local feature-based methods seldom consider the interconnection between local features of vehicles,resulting in the model’s insufficient ability to extract details and difficulty in distinguishing similar samples.2.the traditional attention-based methods ignore the correlation role of different feature channels,which will have certain feature redundancy,leading to poor feature characterization ability and affecting the results of vehicle re-identification.To address problem 1,this paper designs two types of deep learning networks based on local features.Different from the commonly used local feature extraction networks,this paper attempts to propose an LSTM-based local feature extraction network and a graph convolution-based local feature extraction network from two perspectives: the local sequence structure of the image as well as the graph structure.The former can use the memory and forgetting property of LSTM to model the local features of an image and establish the dependency on local features.The latter can use graph structure for information fusion among local features and extract spatially structured features of images.Finally,these two networks are fused with the global feature extraction branch designed in this paper respectively.The rationality and the effectiveness of these two local and global feature fusion structures are verified on two publicly available datasets.To address problem 2,this paper proposes a Channel Correlation-based Attention Model(CCSAM)and embeds it into the global feature extraction branch.By constructing a channel relevance matrix,the feature map of each channel aggregates features with high relevance,thus improving the characterization of global features.Through comparison experiments with other mainstream attention modules on two publicly available datasets,it is shown that the method in this paper can capture better global features,filter the interference of the background and improve the re-recognition accuracy.Finally,this paper uses the proposed algorithm combined with target detection technology to build a deep learning-based intelligent vehicle re-recognition system,which achieves vehicle target detection,specified vehicle re-recognition and trajectory mapping of traffic videos and vehicle re-recognition between cross-camera videos,providing a reference for the development of modern intelligent transportation systems with certain engineering value.