| In our country’s current increasing situation of aging,“Smart Home Care” is one of the important ways to alleviate the pressure of nursing services in our country,which aims to use the current multiple modern technological means,perceive the activity information of the home aged people,and provide timely and remote access to all kinds of nursing care services to the home aged people with the help of Internet and other technologies.This thesis focuses on the problem of indoor human target activity identification in the context of smart home nursing,based on mmwave radar platform,and studies on the methodology of daily activity recognition and sleep activity recognition of human have been carried out.In the recognition of daily activities,the main work for the recognition of short-term and long-term daily activities of indoor human is as follows:1.An integrated daily activity recognition method based on point cloud digital fea-tures is proposed,which uses short-term and long-term daily activity point cloud digital features to train short-term and long-term daily activities and comprehensive decision tree models respectively,and integrates them in Bagging mode to realize the continuous recog-nition of daily activities of the elderly at home.2.Subnetworks for short-term and long-term daily activities recognition based on multidomain feature fusion and point cloud digital features are proposed,and they are integrated by integrated decision tree in Bagging mode,which not only enables continuous recognition of short-term and long-term daily activities,but also enriches the identifiable types of daily activities.In the recognition of sleep activity,the main work for the recognition of human sleep activity under any indoor layout is as follows:1.Successively proposed bed positioning method based on two-dimensional statisti-cal histogram of point cloud center point,recognition method of human state in bed based on statistical deviation and decision rule of point cloud,and error correction method of on-bed state based on static clutter suppression of weak targets,which respectively achieved bed positioning,human target existence recognition in sleep area,and error correction of existence recognition results,and completed accurate and stable recognition of human state in bed.2.A sleep activity recognition method based on point cloud state transition rules is proposed.This method incorporates all the cases of regional state transition into a set of judgment rules according to the law of change of the center point of the target point cloud in three delimited areas or regions during the occurrence of each sleep activity.This method can realize accurate and stable recognition of multiple sleep activities.It also has some anti-interference and practical engineering value.3.A sleep activity recognition method based on MobileNet multi-domain feature fusion is proposed.This method uses several independent MobileNet networks to extract features from each spectrum separately,and then fuses the feature extraction results by channel fusion to the output layer to get the recognition results of various sleep activities.The above indoor human activity recognition methods are all validated by real-time data,which proves the validity of the proposed methods for recognition of daily activities at multiple time scales and sleep activity under any indoor layout. |
