The Algorithmic Research On Person Re-Identification Based On Dictionary Learning

With the continuous improvement of artificial intelligence technology and the accelerated implementation of peaceful city construction,video surveillance system has been rapidly popularized.Cameras everywhere provide a lot of video data for pedestrian identification and tracking.At the same time,it becomes unrealistic to monitor and analyze massive video data manually.The first reason is that the manual speed and the efficiency is low.The second reason is that it will consume a lot of manpower,material and financial resources.Therefore,there is an urgent need for intelligent video analysis techniques such as person re-identification.Person re-identification refers to the technology of judging whether a pedestrian exists in surveillance video by using computer vision technology in the case of cross-device when a specific pedestrian image is given.Because of the low resolution of pedestrian images,inconsistent shooting angles,large changes in illumination conditions,noisy shooting environment,and various pedestrian postures,pedestrian re-identification has become a research hotspot in the field of computer vision,which is both valuable and challenging.The research of person reidentification is generally divided into two directions: one is to construct a distinguishing feature vector;the other is to learn a robust metric learning model.In this thesis,we design an effective person re-recognition algorithm based on both aspects.At the same time,deep learning and feature fusion technology are considered.For different person recognition algorithms,this thesis designs different image feature representation,which can be summarized as follows:(1)Appearance-based superpixel features: because the existing pedestrian image cutting method use pre-set size,it cannot retain the original salient characteristics of the image.In this thesis,simple linear iterative cluster(SLIC)algorithm is used to segment image adaptively.After segmentation,color and texture features are extracted from hyper-pixel blocks with high saliency to construct image feature vectors.(2)Hierarchical representation of pedestrian images: Because there are different global and local information in the image,different information not only trains different models,but also contributes differently to matching accuracy.Therefore,this thesis designs image-level appearance features and three patchlevel appearance features representing different human parts.(3)Although the appearance features can distinguish different pedestrian images,it is difficult to overcome the noise effectively caused by objective factors such as shooting angle,occlusion and so on.Therefore,in order to complement the representation features,this thesis also extracts the deep features based on the neural network.Metric learning has been one of the important directions of person re-identification.Based on dictionary learning,this thesis optimizes its application in pedestrian recognition in different ways.Inspired by the idea of top-push,we proposes a dictionary learning constrained by minimum negative sample(termed as DL-cMN).This method minimizes the intra-class distance and enlarges the inter-class distance by constraining the coefficient coding of minimum negative sample,thus learning a stable and distinctive dictionary model.In addition,in the second person re-identification algorithm,the dictionary learning is further improved,and a joint learning based on dictionary and projection matrix(termed as JL-DPM)is proposed.Based on DL-cMN,joint learning algorithm adds constraints and parameters.Through iterative optimization,sparse coding features and projection matrices can be learned simultaneously,which effectively improves the matching accuracy of pedestrian images.In addition,the feature fusion mentioned in most literatures is only to connect the features in series,without considering the correlation between the features.Therefore,we introduces the probabilistic hypergraph ranking algorithm and generalized discriminant multi-set canonical correlation analysis to fuse different types of features,so as to obtain better sample features.Relationships between multi pedestrians are introduced by building a super-edge,and a multi-hypergraph model is established to fuse deep feature and sparse coding feature.Finally,the similarity between images is calculated by using the probabilistic hypergraph ranking algorithm.Generalized Discriminant Multiple Set Canonical Correlation Analysis use classification information to learn distinctive fusion projection rules,which makes the fused features more conducive to pedestrian reidentification tasks.Finally,in order to verify the effectiveness of the proposed algorithm,we conducted experiments on three standard datasets of VIPeR,PRID450 S and CUHK01 respectively.The experimental results show that the proposed algorithms can achieve good results in person re-identification tasks.