Improved Bof Feature Extracted Algorithm For Pedestrian Re-identification

With the increasing development of social economy and the advancement of safety sense, a large number of surveillance cameras are installed in public places for routine security. Manual labor is unable to handle in face of mass video surveillance data. Therefore, using computer technology for automatic identification and processing of video data is necessary. As we all know, human are the most important objects in the surveillance video. The definition of pedestrian re-identification is whether the target is the same as the real person in different locations and different time, when illuminations, the surveillance and postures of pedestrians would change. It will be of important significance to many fields including criminal case investigation cracks, finding lost children in public places, pedestrians retrieval, gallery management, multi-camera tracking and pedestrian behavior analysis, e-commerce and so on.Pedestrian re-identification is one of the most challenging problems in the image processing and computer vision. Two factors can prove that pedestrian re-identification is challenging. The first reason is that processes are very complex. The overall process of pedestrian is a foreground extraction which is treated to separate pedestrians from the background image. Some inherent features, which are extracted(such as color, texture or super-pixel feature), will be based on an effective fusion. Finally, it will identify and match the feature by similarity measure or other pedestrian re-identification methods. Any further problems will affect the final recognition result. Secondly, pedestrian re-identification problems are facing many challenges such as the impact of the camera, the impact of pedestrians, the impact of video image transmission process and the environment, illumination change and other factors.In this paper, we put forward the improved BOF(Bag of Feature) feature algorithm and fusion covariance descriptors which were used in pedestrian re-identification problems. The main work of this paper is as follows:Firstly, we used the SURF(Speeded Up Robust Feature) to improve the BOF algorithm, because the traditional BOF algorithm has some shortcomings. SURF algorithm extracts the preliminary feature descriptor and generates a visual dictionary in face of invariant by illumination and scale. It can reduce the complexity of the algorithm, improve operational efficiency and solve the problem of high storage requirements by PCA. We combined SURF feature extraction, K-Means clustering and divided the layer of images. The spatial pyramid matching principle is applied to the image histogram's codebook which can make the full use of spatial information of the image and improve the accuracy of the classification.Secondly, we proposed the improved BOF algorithm by LIBSVM and designed an effective classifier. The robust features were extracted by the improved BOF algorithm and the efficient LIBSVM classifier was used into the pedestrian re-identification which can improve the efficiency of the pedestrian re-identification algorithm.Thirdly, we put the covariance descriptor to the pedestrian re-identification algorithm, and use SURF algorithm and spatial pyramid matching principle to replace color, gradients and other features. We not only retained the advantages of improved BOF algorithms for feature extraction, but also joined the high robustness advantages of covariance descriptor. And then we used LIBSVM to match and verify the algorithm. For a small sample of pedestrian re-identification problems which improved the accuracy of the matching.Finally, in this paper, we used the VIPe R dataset, CUHK01 dataset and the ETHZ dataset to compare with SDALF, ELF, LMNN, Bi Cov which were the mainstream algorithms recently. By contrast CMC curve proves this paper's method has obvious advantages over the current mainstream pedestrian re-identification algorithms.