Research On Vehicle Re-identification Based On Deep Learning

Vehicle re-identification technology can find the pictures of a specified vehicle appearing in all surveillance cameras in the real traffic monitoring scenario,which is of great significance to vehicle control in Intelligent Transport System,thus attracting the attention of many scholars.With the wide application of deep learning in recent years,most of the current vehicle re-identification algorithms are based on convolutional neural networks.However,since the convolutional neural network itself has two limitations: the limited receptive field and the loss of detailed information caused by the downsampling operation,the convolutional neural network is not the most suitable model for the vehicle reidentification task.When facing different vehicles with similar appearance,the current algorithm model cannot effectively use the subtle appearance differences between vehicles to distinguish.In view of the above problems,the main research contents and contributions of this paper are as follows:(1)In view of the two limitations of the convolutional neural network,it is proposed to abandon the convolution and use the Transformer,which is currently dominant in the field of natural language processing,to design the vehicle re-identification network model as the benchmark network in this paper.Compared with CNN,Transformer’s self-attention mechanism is not limited by local interactions,and can not only capture long-range dependencies but also compute in parallel.In addition,a vehicle re-identification network model based on Res Net-50 is also designed as a representative of the method based on convolutional neural network,and the experimental comparison is carried out.Experimental results show that the Transformer-based model can outperform the residual network-based vehicle re-identification model.(2)Aiming at the problem that the appearance of different vehicles is almost the same,it is proposed to use the improved YOLOv5 to detect the regional positions of vehicle parts.In order to make the model pay more attention to small target objects,a convolutional attention module is added.In order to enhance the generalization ability of the model,the mosaic data enhancement method of YOLOv5 is improved,and finally the activation function is changed to self-gating Activation function Swish.Experimental results show that the improved YOLOv5 model can accurately check the location of vehicle parts regions.(3)Study the global fusion of local features to complete the task of vehicle reidentification.A local feature branch is added,and the YOLOv5 target detection model is used to detect the position of parts in the output vehicle image,and then the tokens encoded by the image blocks corresponding to these part position regions are also used to supervise the training of the network.Finally,when the model encounters a situation where the appearance of vehicles is almost identical,it can also notice their subtle visual differences,so as to successfully complete the vehicle re-identification task.In addition,in order to alleviate the problem of vehicle cross-view bias,non-visual information,that is,the camera’s view and ID,is also encoded and input into the model,so that the model can learn more robust features.Experiments are carried out on the public vehicle re-identification datasets Ve Ri-776 and Vehicle ID,and the experimental results show that the accuracy and robustness of the model are improved.