Research On Vehicle Reid Algorithm Based On Multi Branch Network

With the development of economy and technology,there are more and more vehicles and pedestrians on the road.It is more and more difficult to control and supervise the traffic only by manual work,which means that it is imminent to take a more rapid and scientific management.As one of the key technologies to achieve this goal,Vehicle Re-Identification can analyze vehicle images using computer vision technology,which has aroused extensive interest in the industry.But up to now,most methods in the field of Vehicle Re-Identification only use global features.Due to the similarity of vehicle appearance,the use of global features alone may lead to the bottleneck problem of accuracy.Few methods simply combine global features with local features,but the local feature acquisition method is relatively simple,which is only limited to the direct segmentation of local part images for feature extraction,and then the image of this area is input directly into the network.What can not be ignored is that these methods only improve the effect of a small amount,and also put forward a greater challenge to the computational power at the same time.In order to solve these problems,this paper constructs a homologous multi branch Vehicle Re-Identification network based on vehicle images captured by traffic surveillance camera.In a novel way,the effective information of global features and multiple local features is combined,and the importance of different local feature branches in the whole feature extraction process is verified,which improves the accuracy and versatility of vehicle recognition algorithm.In the training stage,the local feature branches and global feature branch can be trained at the same time with a small amount of computation only by obtaining the coordinates of the local region corresponding to the image through the third-party target detection framework.In the inference stage,only global feature branch is needed without the assistance of local feature branches to achieve the effect of combining global feature branch with local feature branches at the same time.Inspired by Dropout in Deep Learning,a series of feature fusion methods based on branch Dropout are proposed.But different from the common Dropout,this method realizes the new Dropout between different branches or batches by "Batch Processing" the samples,and selects the training samples through the improved loss function.In order to verify the effectiveness of the proposed method,experiments are carried out using Vehicle ID and Ve Ri-776 as datasets.Compared with baseline,our method has improved by 4.17% on Ve Ri-776 and 6.13% on Vehicle ID.The experimental results show that the improved system architecture can effectively improve the Vehicle Re-Identification results.