A Comprehensive Evaluation of Feature Selection for Gait Recognition Using Smartwatches

Activity recognition that recognises who a user is by what they are doing at a specific point of time is attracting an enormous amount of attention. Whilst previous research in activity recognition has focused on wearable dedicated sensors (body worn sensors) or using a smartphone’s sensors (e.g. accelerometer and gyroscope), little attention is given to the use of wearable devices – which tend to be sensor-rich highly personal technologies. This paper presents a thorough analysis of the current state of the art in transparent and continuous authentication using acceleration and gyroscope sensors and an advanced feature selection approach to select the optimal features for each user. Two experiments are conducted; the first experiment used all the extracted features (i.e., 143 unique features) while (for comparison) a more selective set of only 30 features are used in the second experiment. The best results of the first experiment are average Euclidean distance scores of 0.55 and 1.41 for users’ intra acceleration and gyroscope signals respectively and 3.33 and 5.85 for users’ inter acceleration and gyroscope activities accordingly- providing sufficient disparity in distance to suggest a strong classification performance. In comparison, the second experiment demonstrated stronger results when evaluated (at best the average Euclidean distance scores is 0.03 and 0.19 for users’ intra acceleration and gyroscope signals respectively and 1.65 and 1.1 for users’ inter acceleration and gyroscope activities). The findings demonstrate that the technology is sufficiently capable and the nature of the signals captured sufficiently discriminative to be useful in performing activity recognition. Moreover, the proposed feature selection approach could offer better results and reduce the computational overhead on digital devices.