Person Re-identification Based on Feature Fusion

摘要

Person re-identification is a matching task of person images captured from different camera views. In the real scene, This task is extremely challenging due to changes in pedestrian poses, camera angles and lighting. How to extract robust pedestrian features has become a key step in Person re-identification. In this paper, we propose an person re-identification model based on combined visual features. Our features consist of traditional visual features and convolutional neural network (CNN) features. We extract the CNN feature of person image and fuse them with robust Local Maximal Occurrence (LOMO) features. This fused feature has better performance. Before extracting features, we use the Retinex algorithm to preprocess person images. Finally adopt a random sampling softmax loss to effectively train the model. We experimentally show the effectiveness and accuracy of the proposed method on the VIPeR and PRID450s datasets.