A comparative study of the effect of sensor noise on activity recognition models

To provide a better understanding of the relative strengths of Machine Learning based Activity Recognition methods, in this paper we present a comparative analysis of the robustness of three popular methods with respect to sensor noise. Specifically we evaluate the robustness of Naive Bayes classifier, Support Vector Machine, and Random Forest based activity recognition models in three cases which span sensor errors from dead to poorly calibrated sensors. Test data is partially synthesized from a recently annotated activity recognition corpus which includes both interleaved activities and a range of both temporally long and short activities. Results demonstrate that the relative performance of Support Vector Machine classifiers over Naive Bayes classifiers reduces in noisy sensor conditions, but that overall the Random Forest classifier provides best activity recognition accuracy across all noise conditions synthesized in the corpus. Moreover, we find that activity recognition is equally robust across classification techniques with the relative performance of all models holding up under almost all sensor noise conditions considered.