Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)


Electrical and Computer Engineering (Holcomb Dept. of)

Committee Member

Dr. Adam W. Hoover, Chair

Committee Member

Dr. Elliot Jesch

Committee Member

Dr. Richard E. Groff

Committee Member

Dr. Ian Walker


This dissertation is motivated by improving healthcare through the development of wearable sensors. This work seeks improvement in the evaluation and development of pedometer algorithms, and is composed of two chapters describing the collection of the dataset and describing the im-plementation and evaluation of three previously developed pedometer algorithms on the dataset collected. Our goal is to analyze pedometer algorithms under more natural conditions that occur during daily living where gaits are frequently changing or remain regular for only brief periods of time. We video recorded 30 participants performing 3 activities: walking around a track, walking through a building, and moving around a room. The ground truth time of each step was manu-ally marked in the accelerometer signals through video observation. Collectively 60,853 steps were recorded and annotated. A subclass of steps called shifts were identified as those occurring at the beginning and end of regular strides, during gait changes, and during pivots changing the direction of motion. While shifts comprised only .03% of steps in the regular stride activity, they comprised 10-25% of steps in the semi-regular and unstructured activities. We believe these motions should be identified separately, as they provide different accelerometer signals, and likely result in different amounts of energy expenditure. This dataset will be the first to specifically allow for pedometer algorithms to be evaluated on unstructured gaits that more closely model natural activities. In order to provide pilot evaluation data, a commercial pedometer, the Fitbit Charge 2, and three prior step detection algorithms were analyzed. The Fitbit consistently underestimated the total number of steps taken across each gait type. Because the Fitbit algorithm is proprietary, it could not be reimplemented and examined beyond a raw step count comparison. Three previously published step detection algorithms, however, were implemented and examined in detail on the dataset. The three algorithms are based on three different methods of step detection; peak detection, zero crossing (threshold based), and autocorrelation. The evaluation of these algorithms was performed across 5 dimensions, including algorithm, parameter set, gait type, sensor position, and evaluation metric, which yielded 108 individual measures of accuracy. Accuracy across each of the 5 dimensions were examined individually in order to determine trends. In general, training parameters to this dataset caused a significant accuracy improvement. The most accurate algorithm was dependent on gait type, sensor position, and evaluation metric, indicating no clear “best approach” to step detection. In general, algorithms were most accurate for regular gait and least accurate for unstructured gait. In general, accuracy was higher for hip and ankle worn sensors than it was for wrist worn sensors. Finally, evaluation across running count accuracy (RCA) and step detection accuracy (SDA) revealed similar trends across gait type and sensor position, but each metric indicated a different algorithm with the highest overall accuracy. A classifier was developed to identify gait type in an effort to use this information to improve pedometer accuracy. The classifier’s features are based on the Fast Fourier Transform (FFT) applied to the accelerometer data gathered from each sensor throughout each activity. A peak detector was developed to identify the maximum value of the FFT, the width of the peak yielding the maximum value, and the number of peaks in each FFT. These features were then applied to a Naive Bayes classifier, which correctly identified the gait (regular, semi-regular, or unstructured) with 84% accuracy. A varying algorithm pedometer was then developed which switched between the peak detection, threshold crossing, and autocorrelation based algorithms depending on which algorithm performed best for the sensor location and detected gait type. This process yielded a step detection accuracy of 84%. This was a 3% improvement when compared to the greatest accuracy achieved by the best performing algorithm, the peak detection algorithm. It was also identified that in order to provide quicker real-time transitions between algorithms, the data should be examined in smaller windows. Window sizes of 3, 5, 8, 10, 15, 20, and 30 seconds were tested, and the highest overall accuracy was found for a window size of 5 seconds. These smaller windows of time included behaviors which do not correspond directly with the regular, semi-regular, and unstructured gait activities. Instead, three stride types were identified: steady stride, irregular stride, and idle. These stride types were identified with 82% accuracy. This experiment showed that at an activity level, gait detection can improve pedometer accuracy and indicated that applying the same principles to a smaller window size could allow for more responsive real-time algorithm selection.