Person Re-Identification In The Wild

In the recent years,with the development of deep learning,the tasks including Person Re-Identification in the field of computer vision have made a breakthrough.Person Re-Identification is a subtask of Image Retrieval which searches a most similar person from data base to match the query image.Although face recognition is an important method to recognize person which is widely used in access control,mobile phone unlocking and other fields,it has very strict requirements on illumination and angle,leading to face recognition not very suitable for many scenarios in the security field.Therefore,Person Re-Identification is an necessary research topic.In the natural scene,due to inconsistent camera heights and dense pedestrians,the resolution of person image will be inconsistent and person will be occluded,resulting in reduced recognition accuracy.In this paper,methods such as knowledge distillation,human keypoints detection and graph convolution network will be used to solve the above problems.The main work and innovations are as follows:Firstly,to solve the problem of inconsistent person image resolution in natural scenes,proposing a knowedgedistillation-based multi-resolution OSNet Person Re-Identification algorithm.This method uses OSNet as the backbone network for Person Re-Identification,and uses knowledge distillation to force shallow features to imitate deep features,improve the semantic capabilities of shallow features,and then merge the shallow features with deep features to make the network output pedestrian representations Features with different scales enhance the robustness to pedestrians with different resolutions.Experiments prove that the knowedge-distillation-based multi-resolution OSNet Person Re-Identification algorithm has achieved good results on the Market1501 and CUHK03 datasets.The addition of knowledge distillation has achieved a significant improvement compared to the basic network.Secondly,to solve the problem that the global features of Person Re-Identification have weak discriminativeness,a Person Re-Identification algorithm based on full-scale mining of human keypoints is proposed.The modified method uses the human keypoints detection network to obtain the heatmap corresponding to the keypoints,combines the heatmap with the characteristics of the Person Re-Identification network to obtain the local characteristics of the pedestrian,and uses the global and local characteristics of the pedestrian to characterize the pedestrian.In addition,in order to reduce the negative impact of occluded keypoints on Person Re-Identification,the visibility of each keypoint is predicted to obtain the probability of keypoint visibility,and the probability is combined into the loss function and feature pair distance to reduce the Contribution of keypoints of occlusion.Experiments prove that the Person Re-Identification algorithm based on the full-scale mining of human keypoints has made a significant improvement on the Partial REID Dataset of the occlusion data set.Finally,in view of the second innovation point ignoring the connection between the human keypoints,a Person Re-Identification algorithm based on the graph attention network and the keypoints is proposed.The method mainly adds the GAT module to the Person Re-identification algorithm based on the full-scale mining of human keypoints.This module mainly constructs the connection between the human keypoints and makes the representation ability of the human keypoints feature stronger.This method combines the graph convolution network into the Person Re-Identification network,and uses the graph convolution network to communicate information between keypoint features,thereby improving the characterization ability of local features.In addition,further,this method uses the graph attention network to communicate information between keypoint features.The graph attention network is equivalent to the fusion of the graph convolutional network and the attention mechanism,which can avoid equality when the graph convolution network updates nodes.Look at the problems of all neighbor nodes.Experiments prove that the pedestrian re-recognition algorithm based on graph attention network and human key points has a greater improvement in the Partial REID Dataset than the Person Re-Identification algorithm based on full-scale mining of human keypoints.At last,this article analyzes the shortcomings of the above methods and proposes the follow-up improvement directions.