| With the development of intelligent home space and service robotics, more and moreresearch focuses on the analysis and understanding of service object behaviour accordingto home environment perception and active intelligence service. The key is the analysis ofdaily behavior rules and the judgment of behavior status and needs for service object, thenprovide active service.In different from specific behavior monitoring of fixed spaces, home environment ofdaily life has the characteristics of diversity, heterogeneity and dynamic. The researchresults existed for human behavior mode is difficult to solve the problem of service objectbehavior recognition problem in family environment. In this thesis, according to theservice object in family with normal capacity and living habits, much research work hasbeen done on the behavioral model of global location, feature extraction of action,multi-view action recognition and understanding of the natural behavior.Firstly, the regular pattern of the daily activies of the service object are analyzed, andthen environmental mark points are used in fitting position trajectory in order to establishbehavior model based on global position. At the same time, time stay parametes is addedto optimize the model and verify location behavior of service objects in the homeenvironment, then provides the prediction of active location mode and abnormal judgmentservice.Secondly, based on the characteristics of position and action, the regularity of actiondistribution in different position is discussed. A method of action presention based ondifferent energy image in human pose sequence is given and the local relative twomoments is used to present the characteristics of action, then the combined characteristicsof location and action are clustered through the FCM algorithm based on immunemechanism to obtain the distribution of the object action in different location.Once again, according to the problems of action behavior recognition, arepresentation method for joint calibration and feature extration based on key posture isimplemented. Through circular moment descriptors, the key postures are got from theimage sequence of action behavior and finished point calibration, then the relative weighted joint point of the gesture characteristic is used in action behavior recognitionwith ideal perspective.Then, considering the target actions often have time continuity and the perspective ofspatial continuity, then the space-time probability map can be used as a modeling tool, thespatial distribution of dependence and the adjacent motion with multi-angle are extendedto solve continuous action recognition problem under different perspectives, and themulti-view behavior model of key posture joint based on feature point is the established.Finally, combined with the global position behavior represention and the recognitionof multi-view behavior previously proposed, the high-level behavior in the homeenvironment is analyzed. Fusing time, location, action and physiological parameters, theidentification methods are proposed based on dynamic weights reliability of D-S theory tofurther deepen the service of proactive behavior forecasting and the mechanism ofabnormal behavior judgment. |
