Research On Indoor Daily Behavior Recognition Method Based On Morphological Embedded Features

With the development of society and the improvement of living standards,people pay great attention to safety issues.As an important part of the smart home system,daily behavior recognition has attracted a large number of researchers to invest in this research.Traditional machine learning methods require manual extraction of robust features that are only applicable to specific scenarios and specific behaviors,with significant limitations.The existing deep learning method can learn image features autonomously and identify multiple behaviors in complex scenes.This thesis studies the behavior recognition method based on deep learning,which aims at the daily behavior in indoor scenes,constructs an indoor daily behavior recognition database,and combines the human skeleton and human body parts information to solve the daily behavior recognition problem in indoor scenes.The specific research contents of this thesis are as follows:1.This thesis builds an indoor daily behavior recognition database.The database contains images of different scenes,different resolutions,sharpness,different illumination,different viewing angles,and obvious changes in posture.At the same time,the database will be used as the training data and test data of the behavior recognition method in this thesis,which lays a foundation for the improvement of the accuracy of the method.2.This thesis studies a behavior recognition method based on human skeleton information.In this thesis,a morphological embedded feature extraction network with shallower network layers,fewer parameters and less computation is built,and the human skeleton shape embedded features are extracted to capture the pose information of the object.In this thesis,the idea of feature fusion is used to fuse the original image features and the human body skeleton shape embedding features,and the existing feature fusion methods are improved,so that the fusion features can obtain more powerful expression ability and improve the accuracy of behavior recognition.3.This thesis studies a behavior recognition method based on human body part information.In this thesis,the human body parts are extracted as potential candidate context areas,and the improved multi-instance learning method is used to select the most abundant human body parts morphological features,which makes the behavior classification accuracy rate greatly improved.4.This thesis studies a behavior recognition method based on discriminative features.In this thesis,the features of Convolutional Neural Networks are constrained by intraclass aggregation and inter-class separation,and the discriminative ability of deep learning features is enhanced.Based on the existing intra-class aggregation method,combined with the idea of contrast loss,the improved intra-class aggregation method makes the feature have the characteristics of separation between classes,which effectively enhances the discriminative power of deep learning features and better completes indoor daily behavior recognition task.