Elderly Fall Detection Based On 3D Human Pose Estimation

At the end of the last century,China has entered an aging society.It is estimated that the population of China over 60 will reach 248 million,and the proportion of the elderly population will reach 17.17% by 2020.After entering the aging society,the population of the elderly in China has shown a trend of continuous growth,and aging population will bring many problems and challenges to the society.About one-third of elderly people over the age of 65 will fall at least once a year.More than half of them had various degrees of injuries injured due to the falls.As the age of the elderly increases,the incidence and risk of falls will also increase,even threatening the lives of the elderly.In order to solve the above problems,this paper proposes a fall detection method based on three-dimensional human pose estimation.First one new method for 3D pose estimation integrating 2D pose estimation with large-scale image retrieval.The Human3.6M,HumanEva and CMU MoCap dataset are used to generate our large-scale 2D image database.Each image in this database is closely matched to one 3D pose.The intractable 3D pose estimation from 2D image is then converted into an image matching problem,which is easier in couple with the well-designed efficient image descriptor.Furthermore,an efficient 3D pose optimization algorithm considering the geometric constraints and human body structure is employed on the search results.Second a tree-structured LSTM network for fall detection,taking the 3D human pose as input.This network adopts a method of information bi-directional transmission,using two similar LSTM network structures to transmit information from root to leaf and vice versa.The tree-structured LSTM network uses the temporal relationship between adjacent frames,the relationship of the joints and bi-directional information transmission to improve the accuracy of the fall detection.The 3D pose estimation method proposed in this paper has an average MPJPE of 94.4mm on the Human 3.6M datast.Compared with the state-of-the-art methods,the biggest advantage of our method is that it can be applied in the wild and indoor environment with complex background,not limited to the laboratory environment.The effectiveness of the fall detection algorithm is verified on the fall data set UR-FDD.