The impact of feature vector length on activity recognition accuracy on mobile phone

Abstract

A key challenge for large scale activity recognition on mobile phones is the requirement for producing non-static classifiers that cater for differences in individual user characteristics when performing similar activities in a diverse environment. A static classifier is fixed throughout the system lifetime and does not adapt to different users or environmental changes. Therefore, a personalized recognition model is desirable for each user of the system to ensure accurate recognition in a diverse population of people. One of the main approaches for personalization of activity recognition is the generation of the classification model from user annotated data on mobile itself. However, giving the resource constraints on such devices there is a need to examine the effects of system parameters such as the feature extraction parameter that can affect the performance of the system. Thus, this paper examines the effects of feature vector lengths and varying data set sizes on the classification accuracy of four selected supervised machine learning algorithms running on off the shelf mobile phones. Our results show that out of the three feature vector lengths of 32, 64 and 128 considered, the 128 vector length yields the best accuracy for all the algorithms tested. Also, the time taken to train the algorithms with samples of this length is minimal compare to 64 and 32 feature lengths.