| According to statistics from World Health Organization,falls are the second leading cause of unintentional injury deaths worldwide.In life,the harm of falls to the elderly is immeasurable,especially for the solitary ones.Therefore,it is very important to detect the occurrence of human falls timely and accurately and take corresponding protective measures.Compared with traditional human fall detection technologies,frequency-modulated continuous wave(FMCW)radar becomes one of the research hotspots of fall detection technologies in recent years due to its advantages of non-contact,privacy protection,and unaffected by the environment.However,the existing relevant technologies still have several issues such as a shortage of databases,difficulty in feature extraction,underutilization of features,and low recognition accuracy.To address these problems,this paper builds an experimental platform that uses millimeter-wave radar and investigates technologies for human fall detection.The main work is sketched below:(1)Given the lack of existing public databases and the difficulty of data collection,we built a radar platform to collect data on human falls according to the actual situation,and we constructed a database of human motion,laying the foundation for the follow-up work.After collecting the required data,we adopted a multi-channel high-pass filter to process the echo signals of the radar human body motion,which efficiently suppresses background clutter and environmental noises and improves the signal-to-noise problem of the echo signals.We then performed a time-frequency transform on processed echo signals to obtain the range-time map(RTM),micro-doppler-time map(DTM),and cadence velocity diagram(CVD),which can fully explore the micro-motional trace of the human body and provide rich feature information of technologies for human fall detection.(2)To solve the problem of high computational complexity in principal component analysis(PCA),we introduced two-dimensional principal component analysis(2DPCA)and two-dimensional linear discriminant analysis(2DLDA).Subsequently,we combined each of the analyses with an improved KNN classifier to complete the radar human motion recognition.At the same time,we also proposed a new technology for human fall detection based on 2DPCA and 2DLDA,which has increased the recognition rate of human motions and the corresponding effectiveness of time utilization.(3)To break the limits on information in a single feature,we came up with a smillimeter-wave radar human fall detection method based on the feature fusion sequence network model.The method utilizes fusion technology to reinforce the data domain features,and it also uses a convolutional neural network and a long short-term memory to extract human motion features.The improved networks can increase the recognition rate of human motions. |
