Recognition And Pose Estimation Of Moving Human Body Based On Deep Learning

Moving object recognition and analysis is an important research direction in the field of computer vision and is widely used in our lives,such as smart robots,video surveillance,medical education,sports,national defense,etc.The target recognition process is to select the object of interest in the video or image sequence,find the exact position of the target in the subsequent frame and identify it.Then we can track the selected object accordingly and analyze the meaning of its action to achieve motion prediction or reproduction the corresponding three-dimensional pose effect.This thesis mainly deals with moving human bodies in video or image sequences,such as moving objects in sports events.Through deep learning-based methods,nearlyreal-time moving object detection and three-dimensional pose estimation are achieved.Based on the data collected by common cameras,even with the occlusion and low resolution case,we propose a practical framework named ”DN-2DPN-3DPN”,which can quickly and robustly realize the 3D pose estimation of multi-person motion.Our ”DN-2DPN-3DPN” framework uses three neural networks to deal with the input and is executed in three stages: the first stage is to use the Detect Net(DN)network to detect each person's bounding box separately;the second stage uses the 2DPose Net(2DPN)network to estimate the corresponding two-dimensional pose of each person;the third stage uses the 3DPose Net(3DPN)network to obtain people's 3D poses.The main contributions of this thesis are listed as follows:1.We design Detect Net(DN)based on existing detectors to realize the selection of moving object detector with re-implementation and training.2.We employ different 2D pose detectors and define the corresponding 2DPose Net(2DPN)network for two-dimensional pose prediction mainly based on the result of DN detection.3.We improve the existing network and implement 3DPose Net(3DPN)to achieve prediction from two-dimensional pose to three-dimensional pose,and get better results than existing state-of-the-art methods.Experiments validates the good performance of the system in a multi-user environment and the optimal training method for training our network based on the Human3.6M data set,with real-time multi-person 3D human body pose estimation can be obtained eventually.Based on our ”DN-2DPN-3DPN” framework,the loss accuracy comparison of the benchmark data sets for single-person pose estimation given by experiments demostrate the advancedness and feasibility of our method.We also realizes a practical prototype system for real-time multi-persons motion analysis.