Research On Single Target Cross Lens Pedestrian Recognition Algorithm

With the gradual increase of video data,it is more and more difficult to use the manual screening method to lock and track the target pedestrian to adapt to the large-scale video surveillance system.In the gradual growth of massive video data,how to automatically and quickly retrieve the target pedestrian across the border is an urgent need of the public security.Thesis analyzes and studies the pedestrian cross lens retrieval problem from two aspects.First,in the third chapter,the pedestrian tracking problem under a single camera is studied,and then in the fourth chapter,the pedestrian re recognition problem under cross lens is studied.In the pedestrian tracking direction based on convolutional neural network,the target tracker should not only focus on the region of interest,but also extract more expressive identity features,so as to make the tracker have higher accuracy and better robustness.Based on the existing tracking algorithm(GOTURN),thesis makes the following improvements: By adding attention mechanism module.It makes the feature extraction network pay more attention to the region of interest,so as to achieve the effect of background suppression.Multi level feature fusion.In order to improve the robustness and expressiveness of the features,the cascade method is used to fuse the features of different depths,so as to improve the performance of the target tracking algorithm.Based on L1 norm loss,the loss function of regional overlap rate is added.In the later stage of network training,the loss function of region overlap rate is introduced.The experimental results show that the accuracy of target tracking can be significantly improved.The target tracking method proposed in thesis achieves an accuracy of 49.76 and a robustness of36.15 on the VOT2016 dataset.Compared with the original goturn algorithm,the performance of the method is significantly improved.In the direction recognition of pedestrian based on convolutional neural network,pedestrian appearance attribute is an important semantic information to distinguish pedestrian differences.Pedestrian attribute recognition plays an important role in intelligent video surveillance,which can help us filter and retrieve the target pedestrian quickly.Thesis makes the following improvements based on Res Net50 backbone network:Combine the task of attribute recognition and pedestrian re recognition,improve the accuracy of attribute recognition by improving resnet50 backbone network and dual pooling,and ensure the guiding role of pedestrian re recognition.A parallel spatial channel attention module based on attention mechanism is designed.Based on the two dimensions of space and channel,the specific features of each attribute are located and extracted from the shared feature mapping to improve the expression ability of attribute features.The reasoning fusion module is designed by using graph convolution network to effectively fuse attribute features and pedestrian identity features to obtain more robust and expressive pedestrian identity features,so as to ensure the effect of pedestrian re recognition.The proposed method can achieve 94.74% CMC-1 and 87.02% m AP on the public dataset Market-1501,and 87.03% CMC-1 and 77.11% m AP on the public dataset Duke MTMC-re ID.