Human Localization And Reconstruction Based On Re-identification

With the development of the computer vision,more and more computer vision tasks can be applied to human society.Meanwhile,people have higher and higher requirements for computer vision.The results of some basic computer vision tasks have been greatly improved because of the deep convolutional neural networks.Existing researches of computer vision mostly focus on a single computer vision task and verify the effect of the network on public datasets with proposing some new network structures.The purpose of this article is to integrate a variety of basic vision tasks considering the actual situation in a specific scene and balance the accuracy and time complexity to build advanced vision tasks to meet the more diverse requirements of computer vision tasks in today's society.The main work of this article is as follows:1.Based on the re-identification and target tracking,a human body positioning algorithm is proposed.The algorithm uses a full-scale feature extraction network to extract image features.Besides,it adds a re-identification module based on metric learning,which calculates the similarity of different human detection frames and thereby determines whether different human detection frames are the same person.The algorithm enables precise tracking of specific people in the video.2.Propose a 3D human pose reconstruction algorithm based on deep learning.First,human pose estimation algorithm is improved with proposing a feature pyramid module for extracting multi-scale semantic information and an attention mechanism module for automatically learning semantic information.Secondly,the human body segmentation algorithm is fused as the input of the 3D pose reconstruction algorithm.The 3D pose reconstruction algorithm can reconstruct accurate 3D human models.3.Propose a human motion reconstruction algorithm based on specific scenarios.Firstly,a scene-specific data set is constructed,and a key frame detection network is proposed based on this.The key frame detection network uses a convolutional neural network to extract image features and a long short-term memory module to extract time series relationships.Secondly,an auxiliary object detection algorithm based on object detection detects auxiliary objects to assist the realization of motion reconstruction.The algorithm accurately detects key frames and auxiliary objects in videos.