Study on Fast Human Activity Recognition Based on Optimized Feature Selection

Sensor-based human activity recognition has attracted much scholarly attention due to its merit of wide applicability. However, because of such hardware limitations as battery capacity and computing power, most current generation of wearable devices cannot yet benefit from those activity recognition methods based on deep learning theory and high dimension features since implementing these methods are time-consuming and need a relatively large amount of calculation. A solution, therefore, is proposed for this situation, which aims to reduce computational complexity by reducing the feature dimension through analyzing the linear correlation between the features. Based on the support vector machine model of single-layer fully connected network, the training and recognition time are significantly reduced while the recognition accuracy is still ensured. The experiment is based on the public dataset in the UCI Machine Learning Repository, and it uses Caffe, a deep learning framework, to structure the support vector machine model. In the experiment, when the feature dimension is reduced from 561 to 130, the training time can be reduced by 70% while the recognition accuracy is kept at a promising 91%.