2017, 39(1): 191-197.
doi: 10.11999/JEIT160253
Abstract:
Human activity recognition based on accelerometer embedded in smartphones is wildly applied to inertial positioning, personalized recommendation, daily exercise estimating and other fields. The low recognition rate which caused by varying phone displacement is a crucial problem which needs to solve. To improve the recognition rate when the phones displacement is unfixed, a hierarchical classification-based activity recognition method is proposed. The activity recognition process is divided into multiple layers in this method, and each layer contains a classifier. For training each layers classifier, it runs the feature selection algorithm first, and the classifier is trained based on the selected features. Then, the trained classifier is used to classify the training set, and each samples classification confidence is calculated. Finally, samples whose confidence is lower than the hierarchical threshold are selected as the next layers training set. This process continues until each activitys sample number is less than the predefined pruning threshold. When an unlabeled sample comes, the first layer is used to classify this sample. If the classification confidence is higher than the hierarchical threshold, the recognition is over. Otherwise, the next layer will repeat this process until all the layers are traversed. The experiment collects activity data, and simulates the activity recognition. The simulation show that compared with the current methods, this method may improve the recognition rate from 85.2% to 89.2%.