ABSTRACT
This paper studies methods for identifying human individuals and gender using gait-related features as measured by sensors worn on the body. A recently published dataset due to prior work is used to study the effectiveness of well established and efficient methods for such identifications. The dataset is based on experiments with 16 participants wearing sensors that are part of a widely used gait-sensing platform. The accuracies of the best of these methods compare favorably with those reported by prior work. Since the records in the dataset are characterized by a very large number of fields (323 attributes per record), methods for attribute selection are of particular interest and are also studied. The underlying implementation is briefly described, with a focus on some data management challenges posed by the large number of attributes. A notable result is that prediction accuracies of several competitive methods are not diminished even when the number of attributes is reduced very drastically using attribute-selection methods based on metrics such as ReliefF and Symmetrical Uncertainty.
