Research Of Pedestrian Matching Algorithm Based On Multi-Cameras

With the large-scale application of video surveillance and development of data storage technology,a large number of surveillance cameras appear in a variety of public places,which can monitor people's behavior and ensure our safety.However,attendant massive video data is far beyond the traditional manual video surveillance capabilities.Therefore,video analysis technology is in urgent need of intelligence and automation.Pedestrian matching is an important research direction in the field of video analysis,which mainly solves the problem of judging whether a pedestrian who appears under a camera appears again under another camera,which is so-called pedestrian re-identification under non-overlapping cameras.There are two main directions of pedestrian matching in multicamera: feature representation and metric learning.The pedestrian matching methods generally require many manual intervention,which needs rich experience and a large number of theoretical knowledge to select the most effective and robust feature combination.In addition,most of features are extracted directly from images,or divided horizontally at most,which ignores the characteristics of the human body itself.Besides,most supervised learning requires a large amount of labeled training data,which brings great obstacles to the practical application of pedestrian matching.In order to solve these existing problems,this paper proposes two pedestrian matching algorithms for feature representation.The main contents are as follows:1.In view of the complementarity between local features and the structural characteristics of human body,a pedestrian matching algorithm based on the human structure and feature fusion is proposed in this paper.First,the Custom Pictorial Structure(CPS)is used to detect each part of human body.Then,fused feature of every part is extracted to combine into the final human structure feature.The fused feature includes a variety of color and texture features,such as RGB,HSV,Gabor and so on.Finally,the human structure feature is input into the Relative Distance Comparison(RDC)classifier for training and testing.The datasets used in experiments include VIPe R,i-LIDS and CUHK01,which cover many aspects that need to be tested respectively,such as visual angle,illumination,occlusion and so on.The subdivision of human structure strengthens the feature describe ability,and make the structural characteristics of human body be considered,and can effectively remove the background interference as well.The combination of a variety of color and texture features enhances the robustness of features using the complementary between local features.Good test results have been obtained on three datasets,and the recognition rate of Rank-20 can reach 92.41%,97.01% and 94.45%.2.In view of the excellent performance of the deep learning algorithm in the field of pedestrian matching,a pedestrian matching algorithm based on enhanced deep feature fusion is proposed in the paper to further improve the feature representation.The human structure feature proposed in the previous algorithm,deep feature of Convolutional Neural Network(CNN),and Local Maximal Occurrence(LOMO)feature are combined into an enhanced deep feature in this algorithm.The algorithm has two parts,the first part is to use CNN as a feature extractor to get deep features;the second part is to extract the human structure features.Next,they are combined in the CNN.Then,the feature is connected with LOMO feature into the final enhanced deep feature.Finally,the Mirror KMFA is used to further strengthen the feature,and carry out the metric learning.Through the back propagation of CNN,the fused feature based on human structure detection can dynamically affect the parameters of the convolution layer,thus it can affect the deep feature.The joining of LOMO feature also improves the recognition rate of VIPe R dataset,which has the most severe visual angle change.Moreover,the Max Pooling layer in CNN also has a certain robustness to the visual angle change in some degree,which further improves the results.Good test results have been obtained on three datasets,VIPe R,i-LIDS and CUHK01,and the recognition rate of Rank-20 can reach 97.17%,98.77% and 96.40%.