Research On Behavior Recognition Method For The Intelligent Nursing-care Robot

With the aging of society,lack of human resources and rising costs of labor,intelligent nursing-care robots have a broad development prospect.Behavior recognition technology for nursing objects is a keen research topic of intelligent nursing-care robots.Nowadays,however,less attention is paid to the behavioral patterns of nursing objects and their identification techniques.Therefore,how to identify the behaviors of nursing objects accurately and efficiently is of vital significance to ensure the performance of intelligent nursing-care robots in real scene.Based on this background,this thesis makes study on the behavior recognition method of nursing-care objects.The main contents and innovative achievements are as follows:1.In this thesis,taking the elderly as main research object,video-based behavior recognition scheme is employed to implement the intelligent recognition of behavior models.With lack of a public database at present,a large number of behavior videos photographed on hospital,as a real nursing scene,are used to construct the database.These videos are abundant in imitating the behavior of the elderly,which guarantee the practicability of database.The final database of elderly behaviors includes the training samples and testing samples of behaviors such as walking,getting up,standing and turning over.2.Aiming at the detection accuracy and detection speed of moving object in video,method of moving target detection for the elderly is studied.Combining with discontinuous frame difference and Lucas-Kanade optical flow method,target detection is realized based on the video data organized in this paper.While improve the detection speed,experimental result shows that the proposed method also can achieve better detection accuracy,meanwhile providing basic information for the consecutive behavior identification.3.Aiming at the feature extraction speed and classification accuracy in existing behavior recognition system,this thesis presents a new method.Combining with the characteristics of the elderly behavior patterns,this method applies convolutional neural network to extract behavior feature independently,and integrates with multiple support vector machines to implement multiple behaviors recognition.We design the model in terms of the network structure,activation function,multiple classification mechanism and parameter selection and so on.Compared with convolutional neural network and other recognition algorithms,the cross layer CNN-SVM in this thesis has considerable accuracy and practical significance.