Research On Pedestrian Detection And Re-identification In Visual Surveillance

For video surveillance system,pedestrians are the objects focused.As the scale of video surveillance system incredible growing,it is harder to continuous monitor the pedestrian by traditional means of manual monitor.So the ability of intelligent detecting and re-identification for pedestrian is required by currently intelligent video surveillance system.However,due to the variety of posture and illumination,as well as occlusion and low resolution,making it hard to detecting pedestrian.There are multiple cameras exists in a video surveillance system,and the view angle,posture,illumination,background,occlusion and resolution variety between different cameras,so the appearance large difference between cameras for one pedestrian.While,on the other hand,these factors for one camera is similar for different pedestrians,making it looks similar.All these making pedestrian a challenging problem.This paper focuses on these issues from multiple aspects of pedestrian detection and re-identification in intelligent video surveillance system.First,we studied pedestrian detection in complex background;following the pedestrian re-identification in case of variety viewing angle and posture across different cameras;last,unsupervised metric learning for pedestrian re-identification intensive.The specific contents are as following:(1)Proposed pedestrian detection by block feature contraction,eliminate the influence of background on classification.First,split the sample image of pedestrian into multiple overlapped blocks with the same size,and extraction the shape and texture features,fusion them as the feature of block;then train classifier through the features,remove blocks of poor effects in detection,according to the effect of classifier,these blocks are easily influence by background;then select blocks with good effects,and split pedestrian from background image;finally,shrink features for each block,remove the features which are easily influence by illumination or other factors,and select effective features to forms the feature vector of the block,combine the feature vectors of every blocks for pedestrian detection.Experiment result indicates,the method of block feature contraction,could eliminate the influence of background effectively,and enhanced performance of pedestrian detection.(2)Proposed the pedestrian re-identification method based on part feature importance,solved the disadvantage of single feature could not adapt different appearance.First,split the parts from pedestrian images,and extract color,texture and shape features of each part;then cluster parts with similarity appearance,doing cluster analyze for each parts,following calculating a feature importance weighting vector adapt to all appearance,via error accumulation;last,using this vector to weight feature of parts.Experiment result indicates the influence of view angle,posture and background could be eliminated by part split,and different features could apply on those appearance they adapted,enhanced the pedestrian re-identification performance as well.(3)Proposed an unsupervised metric learning method based on image re-ranking,eliminate the necessary of manual labeling pedestrians across multiple cameras in supervised metric learning.First,calculate initial distance matrix using initial metric function;then re-ranking distance matrix by images re-ranking,and acquire new metric function and distance ranking matrix,while final distance ranking matrix acquired after multiple iterations;last,automatic label positive and negative sample via final distance ranking matrix,doing unsupervised metric learning,and get a more effective Mahalanobis metric function.Experiment result indicates unsupervised metric learning method based on image re-ranking,could eliminate the difficult of labeling massive pedestrian samples,as well as better pedestrian re-identification performance.These three main researches penetrate into pedestrian detection and re-identification application in intelligent video surveillance system,break through the key procedures,and forms the preliminary pedestrian detection and re-identification system for intelligent video surveillance system.Experiments have done on multiple public pedestrian data-set,and result indicates the good effect of this algorithm.This algorithm could be used on field investigation,flow statistics,tracking and so on.