Fall Detection Study Based On Millimeter Wave Radar

For the elderly living alone,falling are the leading cause of injuries among the elderly.So,the timely detection when a fall occurs to an elderly has attracted much research attention.Most of the existing fall detection solutions are based on wearable devices.But the wearable devices often fail to function because most elderly do not wear them.So,for the fall detection of the elderly,the fall detection method based on non-wearable devices is more practical.If not based on wearable devices,the millimeter wave technology has shown significant advantages in body-pose recognition.However,in the past,the existing technology mainly focused on the research of using sophisticated millimeter wave instrument,which complexity makes it difficult to be applied in the scene of monitoring the behavior state of the elderly.In order to take advantage of millimeter wave in the daily life,this thesis studies fall detection based on the ordinary low resolution millimeter wave radar combined with machine learning technology.The main work of this thesis is as follows:(1)Based on the ordinary millimeter wave radar,collected the falling action data,and a data set of falling and non-falling action samples was collected and established for the actual test.The information of the dataset included 9 different action types performed by 46 participants of different body mass index in 4 different environments.(2)Based on the above dataset,the corresponding 3D point cloud is estimated according to the millimeter wave signals reflected from the human body,and four body posture feature extraction algorithms are proposed for feature extraction.Two of them are body height feature extraction algorithms and two are body angle feature extraction algorithms.By the measurement experiments comparison,the best body feature extraction algorithm was selected to extract the eigenvalues of the body height and angle.(3)Based on the above eigenvalues,this thesis designs a fall detection model based on threshold value to judge the fall action of samples.The model is robust to the deviation of human pose feature parameters extracted from millimeter wave signals.The experimental results show that the accuracy of the proposed method can reach 88.6%.