Video-based Person Re-identification

Person re-identification has received extensive attention in academic and industrial circles in recent years.This research mainly includes image-based person re-identification and video-based person re-identification.The main problem it solves is whether the pedestrians are the same person under multiple cameras.This research is particularly important for public safety issues such as finding lost children in shopping malls,or for police officers to detect criminals.However,due to the influence of illumination,occlusion,blur,pedestrian posture and camera angle,the same pedestrian has great changes under different cameras.The video sequences are used to solve the problems in this thesis.Everyone is a sequence under the camera in video-based person re-identification.More abundant spatio-temporal information can be obtained by using video sequence.First,feature representation is considered.Feature representation is to distinguish different pedestrians by learning distinctive features.Unlike static images,which have only spatial information,video sequences also have temporal information.Pose estimation is used to obtain the body joints of the pedestrian,and the high quality walking cycle of the pedestrian can be reconstructed through the ankle points.At the same time,in order to obtain spatial information,body joints are used to divide pedestrians into local regions and extract local features.Finally,all image features of the walking period are connected into the final spatio-temporal features.The final experimental results show that such feature representation can achieve a good effect on the video sequence.Second,metric learning is considered.Metric learning is learning discriminant metric matrix to distinguish different pedestrians.After extracting the spatial and temporal feature of the video sequence,pose constraints are added into the metric learning,which considering that different poses have different temporal information.According to the extracted features,the corresponding poses should be the same.In the final similarity calculation,only the corresponding poses distances are calculated and summed.At the same time,in order to prove the rationality of the algorithm,Dynamic Time Warping algorithm is used to calculate the distance between the two sequences and obtain the warping path.The experimental results prove that the method is the most effective non-deep learning method at present.Finally,deep learning methods are considered for video-based person re-identification.It is two common methods to extract temporal information by using temporal model and predict the qualities of images by using attention model in the field of video-based person re-identification.The deep network based on temporal attention model is put forward by combining these two methods in this thesis.The temporal model is put forward firstly,which the previous frames information are used to improve the quality of the current frame.Then the attention model is put forward to obtain each frame quality score in the sequence,the final feature representation of each sequence is obtained by weighted fusion.The experimental results show that the proposed temporal attention model can achieve excellent performance in video-based person re-identification.This thesis mainly focuses on the research of video-based person re-identification.By using more abundant information of videos,the accuracy of person re-identification is improved.The results of this research are of great significance for public safety.