Research Of Object Re-identification Based On Deep Learning

In recent years,with the popularization of high-definition cameras and the rapid development of computer vision technology,the intelligent surveillance video network has been widely applied in various industries.Person re-identification,that is,non-overlapped multi-camera target matching,is a core technology in the field of target tracking and retrieval and has a great application prospect.The real surveillance video has low image resolution,different lighting conditions under different surveillance scenes or different shooting times,changes in shooting angles,changes in pedestrian posture,etc.The same pedestrian captured by different surveillance cameras could appear quite different,which makes person re-identification more challenging.This thesis makes the following innovations for video person re-identification and image person re-identification:1.In the video-based person re-identification,a background separation technique based on the detection of key points of the human body is adopted in this paper for the background interference that it encounters.First of all,the key point detection technology is used to construct the pedestrian body template for the sample,so as to separate out the background area and acquire the pedestrian gait information on a series of body templates.Secondly,this paper adopts two-stream CNN network for deep feature fusion,effectively utilizes single-step optical flow information and foreground segmentation template method to generate pedestrian gait information and color appearance information,and finally uses RNN network to integrate the information of each step.This judges the similarity between different pedestrian videos.2.In the image-based person re-identification,similarly,we use the human body key point detection technology to construct a pedestrian body template.Different from video-based ones,we use this template as weak supervision information for foreground segmentation,and use it to constrain the neural network gaze area to distinguish the pedestrian's body area from the background area.Specifically,first,we select the feature pyramid network as the network framework of the training phase,and output the foreground prediction models of different scales corresponding to the regions of interest of different network layers in the feature learning phase;secondly,use the pedestrian humanoid template obtained as the preceding The weak supervision information optimizes the prediction model so that the network can better distinguish the pedestrian area from the background area in each stage of feature expression;in the test stage,only the backbone network features of the feature pyramid are used to complete the re-identification work.