Research On Pedestrian Detection Algorithm In Non-overlapping Domain

With the wide application of multi-camera monitoring system,intelligent video retrieval targeting specific targets(mainly humans)has become an important issue in video surveillance.Non-overlapping domain person re-identification is one of the most challenging issues in the field of intelligent video retrieval.It describes that in the non-overlapping multi-camera monitoring system,through a series of image processing technology,to determine if the person appearing in one of the cameras appears in the other cameras.Because non-overlapping person re-identification faces the challenge of some problems such as the low resolution of the target image,the occlusion of the object,the changes of lighting conditions,the change of the person attitude and the different camera parameters,there is still a great distance from the practical application.This paper focus on person target detection and person re-identification.For pedestrian target detection,some existing target detection algorithms are used in the static background and the dynamic background respectively to carry out experiments,and the comparative analysis of various detection methods is carried out according to the experimental results.For person re-identification,a method based on DCGAN is proposed and optimized in this paper.This method uses DCGAN network to extract the characteristics of person images and then calculates the similarity between two images.This method is different from most current person re-identification based on deep learning.The deep learning network is regarded as a feature extractor,which can makethe neural network can use larger data set in the training process,and this method is a semi-supervised learning method for the training process of discriminator D.The experimental results on datasets show that the proposed method has a certain effect on non-overlapping person re-identification.It can also eliminate the negative effects of sharpness,light and occlusion,and has a certain Robustness.Through the optimization of this method,we can also see that the performance of this algorithm can be further improved by feature fusion.