Research On Recognition Of Depression Based On Limb Movements

Depression is a serious mental disorder that is common throughout the world and is characterized by a continuous and prolonged decline in mood,depression,mental retardation,cognitive impairment,impaired physical activity,etc.The existing depression diagnosis process is complex and relatively subjective,and objectively effective auxiliary diagnosis methods are still to be explored.Due to the long-term effects of depression,the clinical manifestations of the patient’s limb movements are characterized by delayed response,motion delay,reduced body coordination,and so on.This thesis is mainly based on human skeletal data and RGB video data on body activity information for methods of depression identification,mainly working for:(1)Identification of depression based on movement and skeletal data and reaction characteristics.The experiment used Kinect V2 devices to record simple sports skeletal data on 25 joint points in depressed and non-depressed patients.Extract the presented time and space characteristics and lower characteristics directly from the original Kinect-3D coordinates recorded.For symptoms of delayed reaction,motion delay,and weakened coordination of physical activity in patients with depression,data pre-processing is carried out based on the timing information of the movement of the skeleton,in conjunction with the field audio information extracted,and the action reaction time of the sample is characterized.This feature was added to the Transformers model structure to assist in the identification and classification of the model,thereby enabling deep learning models to better recognize depression recognition,The new Transformer-React model achieved 72.97 percent accuracy.(2)Depression identification based on action video data.The experiment extracted the stored RGB video data from the original data recorded by the Kinect V2 device and captured information about its body movements.Locate the action clips in the video by calculating the average and standard difference of the sample’s action time and determine the range of video clips to be extracted.By analyzing the action video characteristics,the time-spatial dimensional characteristics in the video data are eventually extracted from the 3D convolutional neural network(3DCNN)model structure,and the sequence characteristics are entered into the encoder structure of the Transformer model to synthesize sequence information through the multi-headed attention mechanism.Compared to other deep learning models,the 3DCNN-Transformer model in this thesis achieved improvements in classification,reaching an identification accuracy of 82.93%.This thesis expects to explore scientific methods of depression identification by preprocessing and constructing identification models of both motion skeletal and video data from extracted depression patients and non-depression patients to assist psychiatrists in objectively and accurately testing depression.