Research On Recognition Of People With Depression Using Skeleton Information

Due to the shortage of medical resources and insufficient personal attention to mental health,various high-risk psychological diseases continue to emerge,with depression being the most prominent.At present,there are many problems with the detection methods for depression,either based on artificial comprehensive judgment,which is inefficient and costly,or based on various invasive biological signal detection,which can bring discomfort to patients.In contrast,skeletal data in gait information has gradually attracted the attention of researchers due to its simple collection methods,non-invasive,low-cost,and difficult to imitate characteristics.In order to solve the problem of insufficient skeleton data in the current depression detection,this study constructed a skeleton dataset of 147 healthy people and 151 patients with depression.Compared with previous studies,the age range has also been expanded,and program optimization has been done for data filtering,denoising and other preprocessing work,providing reliable data support for subsequent experiments.At the same time,three methods for identifying depression populations based on skeleton information have been proposed to address issues such as single representation of skeleton data,weak correlation between joints,and poor performance of feature extraction methods.These methods overcome the limitations of traditional methods and have high accuracy and reliability.(1)A multi-spatial feature fusion method was proposed.This model starts from both spatial and temporal perspectives and processes them step by step.Firstly,the joint position data,joint distance data,and whole bone data within the frame are extracted using Convolutional Neural Network for spatial features.Then,the spatial features are fused and input into Independently Recurrent Neural Network for temporal feature extraction.Finally,the data is inputted into a classifier,achieving a classification accuracy of 69%.(2)A multi-stream spatio-temporal pseudo image method was proposed,aiming at simultaneously extracting local temporal and spatial features of skeleton data.This method constructs two types of skeleton data pseudo images containing spatio-temporal information,one is arranged according to body parts,and the other is constructed based on depth traversal of joints.Two types of pseudo images were input into a self-built Convolutional Neural Network for spatio-temporal feature extraction.After late-fusion,better results were achieved than the multi-spatial feature fusion method.(3)A multi-stream spatio-temporal graph convolution method is proposed,which optimizes recognition performance by changing the structure of skeleton data and increasing the representation form of skeleton information..In order to better adapt the data structure to the skeleton features,this study treats the skeleton data as a graph structure and constructs three skeleton topology structures: natural connection of the human body,connection of the upper and lower limb center of gravity,and connection of the whole body center of gravity,in order to enhance the correlation between the local parts of the skeleton.At the same time,this study constructed an information representation system that includes joint position data,bone position data,joint motion data,and bone motion data to enhance the representation ability of skeletal information.By inputting multi-stream data into the Spatial Temporal Graph Convolutional Networks under multiple topological structures and performing late-fusion,a classification accuracy of up to 77.56% was achieved.These methods overcome the limitations of traditional methods and have high accuracy and reliability.Through the gradual exploration of classification methods for skeleton data,this article found that the local spatial and temporal information of skeleton data is crucial for the recognition of depression,and the degree of tightness between local connections also has a certain impact on the recognition effect,and the fusion method of multiple data streams significantly improves the experimental effect.These findings provide new ideas for depression recognition research based on skeleton information and important references for future research.