Studies On Human Dynamic Stability Analysis Based On 3D Pose Estimation

In essence,instability phenomenon includes the imbalance caused by the center of mass exceeding base of support,the weightlessness caused by the failure to adjust the body posture in time,and the collision caused by the moment of falling.Elderly people are prone to serious consequences,such as fracture,secondary impairment,dysfunction,psychological disorders,which damaging the quality of life and health.With China’s rapidly aging population,fall is a major health problem,which will cause costly for medical system and societies.The theory of human dynamic stability control are of great significance for the prevention of falls in the elderly,the objective estimation of motion diseases,and the objective diagnosis of neurodegenerative diseases.This is mainly due to the fact that different from general behavior recognition,unstable behavior detection relying on the detection technology of 3D poseestimation,has its characteristics.First,the existing dynamic postural stability analysis is still limited to the sagittal plane and cannot describe the real motion completely.Second,the existing posture estimation methods require patients to wear variety of sensors or optical markers,which may affect motion and cause psychological discomfort.These feature increases the difficulty of unstable behavior detection based on 3D poseestimation.In view of 3D reconstruction in instability detection,combining with China’s normal adult human movement characteristics,to the task of poseestimation and instability recognition,we designed the three-dimensional skeletal model and dynamic stability parameters,to realize the human dynamic stability analysis based on3 D poseestimation.We evaluate our methods on Chinese’ datasets and experiment results show the effectiveness and reliable of our methods.The research effort in this thesis comprises the following parts:Part Ⅰ: To human body system with extremely complex morphology and structure,we simplified the human body structure based on the inverted pendulum model,and derived the dynamic stability model based on newton-euler equation.Then,the motion parameters,such as center of mass,velocity and velocity stability domain,acceleration and acceleration stability domain of the single-leg support phase during walking were calculated.The definitions of dynamic stability parameters were proposed.The clear and unified stability variables could sensitively reflect subtle differences among individuals.Part Ⅱ: Given current problems such as the lack of three-dimensional posture datasets,the high difficulty of collection,and the large workload of data annotation,we constructed new dataset with total of 145 cases,grouped by age,including 24 cases for20-50 years old,32 cases for 50-70 years old,45 cases for 70-80 years old,and 44 cases for 80-100 years old.It covered some daily movements,such as single-leg stand,walk,turn-back while walking,avoid-obstacle while walking,go and down stairs,sit-to-walk,squat-to-walk,etc.More than 1.062 million human body poses were collected from different angles,including rich 2D/3D poseannotations,which used to estimate 3D human poseand also calculated the motion parameters,such as center of mass,velocity and acceleration,and plantar pressure.Part Ⅲ: To solve the problem of insufficient expression ability of single human body fundamentals,we proposed the mapping mechanism of Co M position-velocity,Co M position-acceleration and Bo S to construct time-independent region of stability,analyzed the stability of different human body basic surface during left stand,right stand,left kick,right kick,walk,and sit-to-stand.For left stand,right kick,walk,the normalized Co M initial velocity and absolute value of normalized Co M peak acceleration of young group demonstrated significantly larger than these of elderly group.For right stand,left kick,sit-to-stand,the absolute value of normalized Co M initial velocity and normalized Co M peak acceleration of young group demonstrated significantly larger than these of elderly group.It was indicated that the dynamic stability of human body decreased with the increase of age.In addition,for left stand,right kick,walk,and sit-to-stand,the normalized Co M initial velocity and peak acceleration points of young group located ouside of Ro Sv and Ro Sa boundary significantly smaller than these of elderly group.It was indicated that participants’ momentum ability was controlled differently.Part Ⅳ: Aiming at the dependence of 3D poseestimation on optical sensor information,we proposed markerless 3D poseestimation algorithm,OpenPose combined with 3DposeNet and signal processing techniques,effectively estimating the human motor index.Laboratory-based optical motion capture was a reasonable baseline predictor,while 3D markerless poseestimation networks was close to the ground＿truth statistically significantly.For test-retest,the ICC internal consistency of marker-and markerless-based motion analysis were acceptable（ICC>0.41,p<0.05）.It was indicated that markerless-based motion analysis method was reliable.In addition,in comparison of characteristics between marker-and markerless-based motion analysis,the SDC value was small（0.0400.05）,and ICC value exceeded 0.61,indicating good reliability.The Bland-Altman curve distribution remained consistent with the ICC calculation results.It was indicated that our video analysis could be used as a quantitative assessment of gait outside of a clinic and helpful for predicting abnormal gait or fall risk.The main points of innovation in this thesis are as follows:（1）The applicability of the inverted pendulum model was extended,the region amd quantity of stability were defined,which comprehensively analyzing human dynamic stability during movement.（2）The stability boundaries of different movements on different fundamental planes were divided,which accurately describing the stability on the sagittal coronal plane.（3）The three-dimensional human information are directly obtained from RGB image.Comparing with the output of three-dimensional motion capture system recorded at the same time,the proposed methods obtained consistent results.