Cross-domain Pedestrian Re-identification Method Based On Deep Learning

Pedestrian re-identification is one of the hot issues of computer vision technology.It mainly searches and matches pedestrian targets in images or video sequences which acquired by different cameras.Affected by illumination,complex background,occlusion,and different pedestrian postures,the pedestrian re-recognition algorithm has the problem of poor feature extraction robustness in practical applications,resulting in low recognition accuracy and difficulty in meeting actual application requirements.Therefore,pedestrian re-recognition Received the attention of many scholars.This paper focuses on the cross-domain pedestrian re-recognition method based on deep learning,and conducts research on pedestrian detection,pedestrian re-recognition,and face recognition.The specific work content is as follows:In terms of pedestrian detection,the RetinaNet network is used as the baseline model,and a pedestrian detection method based on the joint improvement of the channel interaction mechanism and soft non-maximum suppression is proposed.First,the channel interaction mechanism ECA module is introduced into the feature extraction network of the RetinaNet model to improve the model’s ability to capture details of channel information.Then,Soft NMS is used to replace the traditional non-maximum suppression,which solves the problem of reduced recall caused by the traditional non-maximum suppression operation.Finally,the improved model is evaluated and tested on the Caltech data set.The experimental results show that compared with the original model,the missed detection rate of the improved model based on the two methods is reduced by 0.78%,the AP50 is increased by 0.5%,and the Recall is increased by 2.4%.In the pedestrian re-identification part,two improved methods are proposed based on the AlignedReID++ model.Method 1: By introducing the instance normalization layer IN in the feature extraction network of the AlignedReID++ model,form the IBN layer together with the original batch normalization layer BN in the model.Solve the problem that the network is not robust to different colors and texture style features.Method 2: Add the feature space grouping module SGE on the basis of Method 1,further improve the model’s sensitivity to feature space location information.Test the improved model of method one and method two on three data sets: Market1501,Duke MTMCRe ID,and CUHK03.On the three data sets,Rank1 hit rates increased by 0.8%,2.9%,and 4.7%,and m AP increased by 2.7%,3%,and 3.2% respectively.In the face recognition part,first,the face detection based on DSFD is realized.And evaluated the algorithm performance in the WIDERFACE data set and real scenarios.Then,by introducing the style feature recalibration module SRM into the Mobile FaceNet face recognition network,improved the robustness of the face recognition model to different styles of face images.Finally,the improved model is evaluated on the three face data sets of LFW,CFP-FF and Age DB.The accuracy of the improved face recognition model on the three data sets is increased by 0.25%,0.16%,and 0.3% respectively.In the realization of pedestrian re-identification software,use the Pytorch deep learning framework and Python to implement the core algorithm.At the same time,use Py Qt5 language to design interface and complete the pedestrian re-recognition software.The software has functions such as pedestrian detection,pedestrian re-identification,peer analysis,face detection and face recognition.