Human Motion Analysis And Gait Recognition Based On Deep Convolutional Neural Network

Artificial intelligence is used to analyze human behavior and action,so that the machine can he state of the human body and the intention of the movement.This research has a wide range of applications such as security monitoring,human-computer interaction,virtual ergonomics and so on,and has been studied widely in recent year.In addition,with the improvement of computing power and emergence of large-scale valuable data on the internet,deep learning especially deep convolutional neural networks(CNN)has been greatly developed.In this paper,we will take advantage of the powerful hierarchical representation ability of CNN and focus on the subject of human behavior analysis and gait recognition.We will focus on the following direcitons:video base human pose estimation,serialized signals based human motion modeling and recognition and serialized signals based human gait recognition.This paper mainly includes the following 3 contents:(1)Video Base Human Pose EstimationIn order to achieve 3D human pose estimation,CNN is applied in solving the parameters of SMPL model and camera.For the problem of irregular 3D Human Body Model,the idea of GAN instead of manual rules is used to constrain the process of generating 3D Human Body Model.For the problem of jitter and unique solution,we apply the constraints of similarity between video sequences,consistency with the initial state and human's height in video based human pose reconstruction.For comparison,a motion capture platform with 2 Kinect(version 2)and ipiSoft series software is build.Experiments are conducted to validate the effect of speed,position,orientation and complexity of actions.The results show that our proposed method achieves the same performance as the commercial motion capture platform.Finally,we develop a human pose estimation based dance grading program to support this application research.(2)Serialized Signals Based Human Motion Modeling And RecognitionCNN is introduced into the modeling and recognition of serialized signal based human motion.As to the 3D human pose data and electroencephalogram(EEG)data with sparse space,a kind of feature map containing time and space information is proposed,which transforms the problem of serialized signals based human motion modeling and recognition into the field of image processing.For human posed based action recognition,we introduce ResNet for action recognition and optimal solution is obtained through experiments.The proposed solution achieves the best performance on NTU RGB+D dataset and SUB Kinect Interaction dataset for action recognition.For action detection,the optimal feature is used as input of fast R-CNN.We replace the region generating network with sequence generating network and region of interest pooling with region of interest alignment.Finally,the proposed method achieves the best performance on PKU-MMD dataset for detection.For EEG signals based human action modeling and recognition,the structure of CNN is optimized according to the characteristics of EEG signals,and the state-of-the-art performance is achieved on BCI Competition(Data Set 1)dataset and the dataset collected by ourselves.(3)Serialized Signals Based Human Gait RecognitionCNN is introduced into serialized signals based human gait recognition.For the problem of views,we propose a CNN based gait representation and introduce Joint Bayesian to model view variance.For both the verification and identification tasks,the proposed solution achieves the best performance on OU-ISIR Large Population Dataset(OULP)and CASIA-B Dataset(CASIA-B)in different multi-view settings(no-view setting,cross-view setting and uncooperative setting).For the condition of non-linear walking,we proposed a method which combines the human pose estimation and CNN.The method is validated on the IPI Dataset collected by ourselves.The results show that the proposed method outputs the human silhouette based method.Additionally,the 3D pose obtained by our proposed method and commercial motion capture platform achieve the similar performance in human gait recognition.