Person Re-identification Based On Convolutional Neural Network And Orientation-Pose Information

Person re-identification is a computer vision task that identifies all images of a target person taken in different cameras,with one query image captured by another camera of the multi-view video surveillance system without cross-view overlapping.As an extremely chanllenging task in the field of intelligent video surveillance,person re-identification has attracted a great deal of attention from computer vision scholars in recent years.In the current public video surveillance environment,the captured images are generally characterized by diverse shotting views,large differences in pose,complex background,significant changes in illumination,and serious occlusion,which bring great difficulty to solving the problem.In the early years,the works on the person re-identification mainly focused on the manual design of the features about color,texture and saliency,as well as the learning of distance metric,while the re-identification methods based on deep learning which has emerged in recent years integrate feature extraction and metric learning into an unified framework to obtain overall optimized re-identification models.In order to extract discriminative features which can overcome the difficulty of matching caused by the appearance difference created by the diverse views and spatial misaligment,this paper proposes a person re-identification method based on convolutional neural network and orientation-pose information.The convolutional neural network has the advantage of automatically learning discriminative image features,while the orientation and pose information plays a role of extracting effective local features.The entire re-identification process includes three steps:(1)In the preprocessing step,a three-channels person orientation recognition network designed by this paper is used to obtain orientation information,which includes four categories: front,back,left and right.At the same time,the pose estimation method Convolutional-Pose-Machine is used to obtain pose information,which is defined as the locations of 14 joints of the human body in the image.(2)In the training step,the raw image and the different local part images located by pose are input into the convolutional neural networks respectively,to train the network models that can extract discriminative global and local features.(3)In the re-identification step,according to the orientations' combination of the image pair,multi-strategy weighting is performed on different local features,then,the global feature are concated to form the final re-identification feature.The cosine distance between different image pairs are calculated,and the matching results are obtained by sorting the distances.In this paper,we design and test experiments on two large public ReID datasets Market-1501 and CUHK03,as well as five traditional smaller datasets including CUHK01,VIPeR,3DpeS,PRID and i-LIDS.Our method achieves higher accuracy on both Market-1501 and CUHK03 datasets,and the ReID re-id effect of our method on traditional smaller datasets also has certain advantages compared to the state-of-the-art algorithms,which proves the effectiveness of our method.