Training Classifiers with Shadow Features for Sensor-Based Human Activity Recognition.

摘要: In this paper, a novel training/testing process for building/using classification model based on human activity recognition (HAR) is proposed. Traditionally, HAR has been accomplished by classifier that learns the activities of person training with skeletal data obtained from motion sensor, such as Microsoft Kinect. These are spatial coordinates (x, y, z) different parts body. The numeric information forms time series, temporal records movement sequences can be used classifier. addition to features describe current positions in data, new called ‘shadow features’ improve supervised learning efficacy Shadow inferred dynamics body movements, and thereby modelling underlying momentum performed activities. They provide extra dimensions characterising process, significantly accuracy. Two cases tested using trained shadow features: one wearable sensor other Kinect-based remote sensor. Our experiments demonstrate advantages method, which will have an impact detection research.