Research On Person Re-identification In Multiple Deep Feature Fusion

Person re-identification(re-id)is an important application of intelligent surveillance system.Re-id aims to search people by matching person images or videos across non-overlapping camera views deployed at different locations.In multi-camera surveillance system,tracking technology help users to determine the identity of targets within the same camera or across overlapping cameras.But when identification come to across non-overlapping cameras,re-id technology can help.In other words,Re-id fills the technical gap of identification in the cross-cameras target tracking system.There are many difficulties to overcome yet,such as occlusion,inconsistent camera views,poor lighting conditions,large background changes and constantly changing pedestrian poses.Re-id has both research value and challenge in the field of computer vision and attracts the attention of researchers all over the world.In this paper,the fusion mechanisms of multi-type pedestrian features in re-id under the framework of deep learning are mainly investigated and analyzed.Furthermore,our method also improve the accuracy of existing re-id networks through multi-type deep feature fusion.The purpose of fusing multiple types of deep features is to overcome current difficulties in re-id.The main achievements contain three aspects as follows:(1)The fusion mechanism of pedestrian appearance and attribute features in the framework of deep learning is investigated and analyzed.In order to solve the problem of insufficient discriminative power of existing re-id network features by fusion attributes,The TriHard loss based multi-task(THM)person re-identification network is proposed in the framework of deep learning.THM can obtain more pedestrian discriminative information by learning identity and attribute margins simultaneously.Firstly,the pre-trained ResNet-50 is loaded to extract pedestrian features of pre-processed images.Secondly,pedestrian features are fed into the designed multi-task network which consists of two branches.The two branches are trained jointly by minimizing combined TriHard loss of identity and attribute.Finally,the trained model is used to extract pedestrian appearances and attributes features.The features are used for person re-identification and attributes recognition.On Market-1501 and DukeMTMC-reID,rank-1/mAP of THM reaches 91.7%/87.9% and 85.4%/81.6%,which shows that features extracted from the proposed network are more discriminative.(2)The fusion mechanism of pedestrian appearance and pose features in the framework of deep learning is investigated and analyzed.Images of a same person have extreme variation due to viewpoint,pose,occlusion and detection error,which are major challenges in person re-identification.For the purpose of improving the accuracy of person re-identification under pedestrian pose changing,a correlation channel-wise based part aligned representations(CCPAR)for person re-identification is proposed.Firstly,person images are input into two subnetworks to obtain person appearance and part features respectively.Secondly,a correlation channel-wise module(CCM)is designed for optimizing the channels weights of part features.CCM mines the correlation between channels of part features.Finally,appearance features and optimized part features are fused by bilinear pooling.Experiments on three large scale datasets and results show that CCM can enhance part features.Experiment on Market-1501,DukeMTMC-reID and CUHK03 datasets,rank-1/mAP of CCPAR reaches 93.9%/90.6%,87.6%/83.3% and 70.4%/72.8%,which is superior to the other existing methods.(3)The fusion mechanism of pedestrian appearance and semantic part features in the framework of deep learning is investigated and analyzed.In order to alleviate the background clutter in pedestrian images,and make the network focus on pedestrian foreground to improves the utilization of human body parts in the foreground.In this part,we propose a person re-identification network that introduces semantic part constraint(SPC).Firstly,the pedestrian image is input into the backbone network and the semantic part segmentation network at the same time,and the pedestrian feature map and the part segmentation label are obtained respectively.Secondly,the part segmentation label and the pedestrian feature maps are merged to obtain the semantic part feature.Thirdly,the pedestrian feature map is obtained and the global average pooling is used to gain global features.Finally,the network is trained using both identity constraint and semantic part constraint.Since the semantic part constraint makes the global features obtain the part information,only the backbone network can be used to extract the features of the pedestrian during the test.Experiment without re-ranking on the Market-1501 and DukeMTMC-reID,Rank-1/mAP of SPC reaches 93.6%/83.6% and 85.4%/71.3% respectively.