Research On Human Behavior Recognition Based On Heterogeneous Information Fusion

HAR(Human Activity Recognition)technology is a key research direction in the field of artificial intelligence.Due to its focus on mining human behavior-al patterns and future trends,its application in health monitoring and safety guidance has become more and more extensive.This article mainly studies the fusion technology of heterogeneous data type,and carried experiments on two designed algorithms in a hierarchical and a structural model,and finally imple-ments an HAR system based on the two collaborative models.Firstly,a hierarchical integrated random forest algorithm is designed for the multimodal data of non-homogeneity.Non homogeneous multimodal data such as Wi-Fi and acceleration,due to their inconsistent data distribution and differ-ent analysis methods,can not be analyzed just by a simple combination.But their combination can be helpful for studying the behavior patterns under certain environments,so researchers have paid much attention on this study.This paper proposed a hierarchical algorithm model,integrating the data fusion and context aware ability on the lighter layer to guide the classification task on a higher layer.This design greatly reduced the stress of HAR algorithm to deal with various kinds of daily behavior.By expanding the advantage of random forest algorithm,this model can recognize 30 kinds of human activities in four occasions with the precision above 95%and can perform HAR tasks on the mobile phone.Secondly,a structural data representation method is adopted for the multi-modal data fusion and a convolutional neural network is implemented based on the designed input data.Convolutional neural network is mainly used for image recognition and classification,with its advanced ability of nonlinear expressive-ness and back propagation in automatic feature extraction and learning tasks.Therefore,compared with the process of manual feature extraction,this network has a stronger generalization ability.This paper experiments on combining mul-timodal data in different channels of an image and input it into the CNN model,designs the inner structure of the network such and convolutional layers and pooling layers,generates an effective automatic learning model.This model can recognize 6 kinds of basic human activities and can perform HAR tasks on the server.Finally,an HAR system is designed based on the collaboration of the two models.The ensemble random forest algorithm has less parameters than the CNN model and require less computational storage,so it is run on the mobile phone and use the convenience of Wi-Fi data to recognize the location before classifying the activity types.The CNN model is run on the server because it has much more parameters to store and require fast computational device,and it is used to carry out offline learning and high computational tasks.The combination of them eventually generates an HAR system that can learn both online and of-fline.