Research On Person Re-Identification Based On Deep Feature And Re-Ranking

Person re-identification(re-id)is a matching task with the aim of retrieving target pedestrian among the gallery set under cross-view condition.For the environmental factors,such as illumination,view,pose and occlusion,the pedestrian appearance from the same ID may show great differences which bring about much difficulty in person re-identification.Our work focuses on two problems of re-id in terms of feature description and re-ranking,then we apply the researches to the re-id system on campus in the real world.The main contents are as follows:(1)For improving the robustness of person's features to illumination and pose variation,on the one hand,this paper first uses the MSRCR algorithm to preprocess the original image which can enrich the local color information by enhancing the darker areas of the image.On the other hand,we utilize the pose estimation algorithm and affine transformation to normalize the pose,and design a deep neural network,PIF to extract the deep features which contain the pose invariant information.Compared with handcrafted features and local features after pose normalization,this model learns original images and normalized images jointly by pseudo-Siamese structure,it takes the both global information and local information into consideration,and improve the robustness of pedestrian features to illumination and pose variation.Finally,we conduct experiments on three challenging datasets and prove the effectiveness of our method.(2)For improving the quality of the person re-identification ranking,we apply the re-ranking algorithm based on context to optimize the original ranking and improve the re-id' s accuracy.Nevertheless,the effect of this kind of algorithm is very sensitive to initial ranking as well as introduce some noise during the re-ranking.To this end,we make some improvements on a re-ranking algorithm based on context information,and propose a bidirectional KNN based re-ranking algorihm.We calculate the Jaccard distance between the sets which satisfying the bidirectional KNN relationship,and aggregate original distance and Jaccard distance to optimize the initial ranking.At last,we demonstrate the effectiveness of our re-ranking algorithm by conduct experiments on two datasets.(3)Based on the researches mentioned above,we apply the re-id technology on campus environment and design a re-id system in the real world.For the original video frame captured by the camera,we first use the object detection algorithm to extract the bounding box of the pedestrian.Next,the pre-processing,feature extraction and similarity measure of the pedestrian's bounding boxes are sequentially performed,and the obtained initial ranking is optimized by re-ranking.Practical application proves that our system can achieve the function of target pedestrian recognition in the campus scene.