Research On Complex Human Activity Recognition Based On Wearable Intelligence

Activity recognition based on wearable devices is widely used in smart-home health care,especially for the elderly.For example,sensors embedded in the soles of the shoes could sound an alarm to prevent falls in the elderly.However,these studies are all carried out in a controlled environment.In practical application,there are many difficulties in recognizing complex activities due to less-attachments,low sensitivity of consumable wearables and subject's variation.So,in this paper,we aiming at the recognition of complex human activities in daily uncontroll environment,and the main work of this paper is as follows:(1)We have verified the limitations of traditional algorithms in the recognition of human activities in uncontroll environments,and compared various algorithms including deep learning and general supervised learning methods to analyze what types of activities the algorithms are suitable for.We also evaluated a typical machine learning method,artificial neural network,from different field,such as the influence of the sliding window's size,the influence of activation function,and the influence analysis of different feature sets.The results show that the ANN model is robust to recognize continuous dynamic activities and static activities,but not to transitional activities.(2)We researched the sensor-based fusion method and the feature fusion method respectively,so as to achieve the high-level recognition effect by using low-level equipment for activities.We compared the data collected by different types of sensors,such as accelerometer,gyroscope and magnetometer.After the fusion of different parts and different types of sensors,the accuracy was improved and the robustness of the model was improved.(3)We design a two-layer hybrid hierarchical model which is defined with components of IightGBM(LGB)and Artificial Neural Networks(ANN)for accessing and evaluating cost-effective wearable intelligence approaches for PA recognition in free-living environments.The hypothesis of this model is first suggested utilizing lessattached on-body consumable wearables like belt and wristband devices,and building up a PA dataset collected in free-living environments like elderly home,hospital,office and gym.Experimental results show that the performance of our model is better than that of traditional learning methods in both controlled and uncontrolled environments.