Abstract: Sensor-based human behavior recognition has been widely used in health monitoring,motion analysis and human-computer interaction.Feature selection acts a critical step when identifying human behaviors accurately,aiming to improve classification performance by selecting classification-related features,so as to reduce feature dimensions and computational complexity.The absence of feature redundancy,nevertheless,poses challenges to legacy feature selection methods.Therefore,to resolve the insufficiency that only feature correlation but not feature redundancy is involved in the Discernibility of Feature Subsets (DFS)-based feature selection method,a novel Redundancy and Discernibility of Feature Subsets (R-DFS)-based feature selection method is proposed to incorporate the redundancy analysis into feature selection process and remove redundant features,so as to improve classification accuracy rate and reduce computational complexity as well.Experimental results reveal that the improved method can efficiently reduce the feature dimension with the improved classification accuracy.

Key words: Acceleration sensor, Behavior recognition, Feature correlation, Feature selection, Feature subset discrimination, redundancy