Research And Implementation Of Video-based Person Re-id Based On Local Representation Learning

With the construction of smart cities and people's attention to safety issues in public places,a large number of surveillance cameras are placed in various public places to ensure our safety.Widely popular surveillance cameras provide massive amounts of video data,and how to quickly and effectively identify specific target characters in these video data becomes a key issue.Person re-identification(re-id)is to identify whether the pedestrians captured by two surveillance cameras that do not overlap in the field of view are the same person.The pedestrians are mainly identified by the comprehensive information of the pedestrians.The pedestrians do not need active cooperation and have great research value.Person re-id mainly includes single-frame image-based person re-id and video-based Person re-id.Compared with single-frame images,the pedestrian image sequence in the video contains richer and more comprehensive pedestrian information.And in some practical situations,video-based person re-id is greater application value and application needs.In order to more accurately compare the overall characteristics of pedestrians in different videos,this paper proposes the research and implementation of video-based person re-id Based on local representation learning.The implementation process includes:first using residual-recurrent neural network(Res-RNN)to extract pedestrian spatio-temporal features,and then using local feature descriptors to characterize pedestrian spatio-temporal features,to obtain automatically aligned pedestrian global features,and calculate the distance between pedestrians in different videos to determine whether they are the same pedestrian.For the extraction of pedestrian spatiotemporal features,this paper uses the convolutional neural network(CNN)part in the classic convolutional neural network-recurrent neural network(CNN-RNN)structure,and uses the Res-RNN proposed in this paper to optimize the recurrent neural network to extract pedestrian spatio-temporal features.The residual-recurrent neural network is a recurrent neural network structure,which forms a residual structure through shortcut connections between the input layer and the hidden layer at different times,and converts the deep recurrent neural network expanded in time series into a collection of shallow networks,optimize the performance of the network and extract the spatio-temporal features that better represent pedestrians.In view of the differences in pedestrian poses in different videos,this paper uses local feature descriptors to characterize pedestrian features and obtain pedestrian features aligned on the overall model.Since the local feature descriptor is a set of local common features learned from network training,and pedestrians in different videos are characterized by the same local common feature set,the automatically aligned pedestrian representation vectors are obtained.This paper conducted a comparative experiment on the commonly used video pedestrian re-recognition dataset PRID 2011 and iLIDS-VID.The experiment shows that the video person re-identification method proposed in this paper has achieved good experimental results and it is easy to implement.