Multi-target Human Behavior Recognition Based On Video

Visual ability has been an important approach to acknowledge the world for most of animals for hundreds millions years.Video-based human physical behavior analyzing is for video-based human physical detecting,tracking,recognizing body behaviors and understanding man-man,manenvironment interacting relationship without more human interfere.Video-based human physical behavior analyzing has its great value in intelligence video supervision,intelligence man-machine interaction and AI fields.This article uses Microsoft Kinect for Windows 2.0 as video source gatherer.Deep researching on human physical character gathering and consecutive behavior disassembling with 3D human spinning information and conclusions are drawn below:(1)Defined key frame distance of deep image which realize the disassembling and recognizing of human physical behavior.Nowadays,physical behavior recognizing is based on assuming that human physical behavior start frame and end frame are set.To this limitation,this article defined key frame distance based on deep image and provided the algorithm of video-based key frame distance of human physical behavior disassembling which make single moving target's consecutive behavior disassembling more simple and convenient.(2)Multi-target human behavior recognizing based on joints quaternion.This article is based on quaternion theory which combined 3D human body skeleton and provided the concept of using human body joints quaternion as character information.It is able to use to build human physical behavior model.Firstly,separate multi-target into individuals.Then separating multi-target human physical behavior from video clips by using SVM as classifier.It is able to realize video-based multi-target human physical behavior recognizing in less character and high accuracy.(3)The abnormal behavior recognizing based on Hidden Markov Model(HMM).This article expands the uses of 3D human spinning information to the abnormal behavior recognizing which uses joint angle as character parameter and combines HMM algorithm to realize the abnormal recognizing of consequence human physical behavior from the video.Also analyzes the effects of abnormal behavior recognizing accuracy that caused by different character information.Finally,summarizes the research content in the paper,points out the shortages existing in the research,and prospects the next research direction.