Research On Key Technologies Of Pedestrian Recognition Based On Non-overlapping Visual Filed

Video surveillance technology is a research hotspot in the field of computer vision.Pedestrian recognition based on non overlapping visual field has always been a research difficulty in the field of surveillance,which is of great significance.Pedestrian re recognition refers to finding and matching the target pedestrian’s moving track under other cameras without cross vision according to the given image of the target pedestrian under the camera.However,in the actual monitoring,there are some influencing factors and challenges such as pedestrian pose change,illumination change,angle change,etc.,which lead to low resolution of pedestrian image,obvious visual blur,and limited effect of measurement learning and recognition.Therefore,this thesis uses equalization and local block methods to preprocess the pedestrian image,and uses texture,color and other feature fusion and multi feature measurement learning methods,mainly for the two important steps of pedestrian recognition technology: feature extraction and similarity measurement.The main work and innovations of this thesis are as follows:(1)This thesis expounds the theory of pedestrian recognition technology,including feature extraction and similarity measurement.Firstly,by introducing many typical pedestrian features and many traditional similarity measurement methods in detail,this thesis sums up some shortcomings and shortcomings of single feature and traditional measurement methods.Then,based on this,Cross-view Quadratic Discriminant Analysis(XQDA)improves the recognition rate of pedestrian recognition.In addition,it summarizes the difficulties and challenges faced by pedestrian recognition technology in the actual monitoring scene,as well as the common application scenarios,which lays the foundation for the later research and improvement of the algorithm.(2)The research of image preprocessing based on minimum equalization is proposed.Firstly,aiming at the disadvantage of poor recognition effect of single feature,this thesis proposes a method to extract part of local block features by using multi feature fusion such as texture and color features;secondly,aiming at the unclear contour of pedestrians,this thesis proposes an image processing method of minimum histogram equalization;and then combines the pedestrian features after minimum equalization with XQDA metric learning to form the pedestrian recognition after minimum equalization Finally,the proposed algorithm is simulated in data sets such as VIPe R,PKU-Reid,i-LIDS-VID and prid＿2011.Among them,the re recognition rate on the data set VIPe R is 38.13%.Compared with the existing algorithm,the proposed minimum equalization pedestrian feature is more effective for re recognition.(3)Based on the preprocessing of the minimum equalization image,an algorithm of multi-model and multi-scale LBP(MM-LBP)feature fusion is proposed to enhance the ability of feature identification.Firstly,the traditional LBP,MB-LBP,circular LBP,rotation invariant LBP and other features are simulated;secondly,the circular LBP features with better effect are fused in multi-scale and multi-model;finally,the distance fusion algorithm with multi features measured separately is proposed.The experimental results show that the proposed algorithm improves the recognition rate.