An Effective Algorithm to Detect Abnormal Step Counting Based on One-Class SVM

Walking is the most easily achievable way to reduce obesity and to keep fit. Step counting devices are helpful for users to develop regular walking habits, but they may mistake some abnormal situations (SHAKING, SWING, JITTER, PENDULUM, etc.) for walking and thus result in step counting errors. Such errors will inflate the number of steps, and lead to biased inter-user comparisons or competitions. To fix the problem, we developed an effective algorithm to detect abnormal step counting, which extract 13 non-directional features from acceleration data, and feed them into One-Class Support Vector Machine (OCSVM) to check wether it is abnormal. The detection algorithm has been tested on a dataset composed of 600 minutes real world data, which is captured by 10 users. It is shown that the proposed algorithm is capable of detecting most of the abnormal situations while not mistaking normal activity (walking/running) for abnormal. While the false alarm rate is 5.08%, the algorithm detects 90.72% of abnormal situations. The AUC (Area Under the Curve) of the proposed algorithm reaches 0.9252, which shows that the algorithm is very effective for abnormal situation detection.