An optimal signal selection method based on feature neighborhood using for human gait mode recognition

Abstract

Signal selection is a crucial issue for classifiers. A feature neighborhood based method is proposed to select the optimal signals in this paper. The neighborhood of a feature is first defined to classify each example according to whether its feature value belongs to the neighborhood. A discernibility matrix is established based on all the classification results of the feature. The classification ability of a feature is then calculated according to its discernibility matrix. Furthermore, the classification ability of a signal combination is achieved by calculating the disjunction of the corresponding feature discernibility matrices. The optimal signal combination (O-combination) with the highest classification ability is finally selected. The proposed method is validated by using the data from two public databases of human motion. The results show that 2∼5 signals are selected from over 50 signals to be the O-combinations for human gait mode recognition. The accuracy rates of gait mode recognition (RARs) of these O-combinations are higher than those of other signal combinations with the same dimensionality. The RARs of the 4- and 5- dimensional O-combinations are even higher than that of the combination formed with all signals. Specifically, the RARs of the 4-dimensional O-combinations are 99.37% and 99.86% on the two public databases. In addition, it is found that the signal of clearance between swing foot and stance foot exists in most of the combinations with high RAR, which indicates that this signal cannot be ignored for gait mode recognition.