| With the improvement of people's safety awareness,video surveillance technology has developed rapidly,and a large amount of monitoring probes have been deployed in the streets to form an increasingly dense monitoring network.The technique of using the sights of neighboring cameras to search out target pedestrians and continue tracking,that is,pedestrian re-identification technology has received wide attention..However,the pedestrian re-recognition technology will be interfered by factors such as lighting conditions,high storage cost,difficulty in image alignment,and various pedestrian attitudes in the video.The different parameters of the monitoring probe will also cause differences in the resolution of the surveillance video,which is not conducive to recognize specific person.Based on these difficulties,the dissertation conducts a study and proposes a method for recognizing pedestrians by combining Fisher Vector(FV)and Convolutional Neural Network(CNN)features.The content of the thesis research has the following points:(1)A person re-identification method that using FV to quantify person features is proposed.Aiming to solve problem that conventional eigenvectors extracted by the person re-identification method need more storage space and the computational efficiency is low,a person re-identification method that using the FV to quantify the extracted features is proposed.This method uses FV to quantify the combining of the color names and color histograms that extracted from the person images,and uses the obtained FV to describe the features of a person,then completes the work of recognizing person by calculating the differences between different person trait descriptions.The experimental results on the Market-1501 dataset show that compared to the use of conventional eigenvectors,this method maintains a low memory footprint and has a higher recognition accuracy in person re-identification.(2)A re-identification method of person image feature description based on weighted combination of FV and CNN features is proposed.Aiming at the problem that the existing mainstream CNN technology is susceptible to image rotation,a person reidentification method based on certain weights combined with CNN features and FV is proposed.This method makes use of the similarity of encoding and aggregation process between CNN and FV,and outputs the training results obtained by the intermediate layer during CNN training,combines this results with the quantified FV by certain weights to obtain a new image feature representation.Finally,the person reidentification is performed by calculating the similarity of different image representations.The experimental results on the Market-1501 dataset show that compared with the technique using only CNN,the image features extracted by the thesis have rotation invariance,higher recognition accuracy,and better compressibility.(3)Design and implementation of person re-identification system.Based on the above research,combined with OpenCV graphics library,VLfeat graphics library and Qt interface library,the pedestrian re-identification system is designed and implemented.The system is divided into an image acquisition module,a data processing module,a feature matching module,and a core pedestrian recognition module,and a user-friendly operation interface is constructed to realize the function of person re-identification. |
