Date of Award
Master of Science (MS)
This thesis explores a machine learning-based approach to step detection and counting for a pedometer. Our novelty is to analyze a window of time containing an arbitrary number of steps, and integrate the detected count using a sliding window technique. We compare the effectiveness of this approach against classic deterministic algorithms. While classic algorithms perform well during regular gait (e.g. walking or running), they can perform significantly worse during semi-regular and irregular gaits that still contribute to a person’s overall step count. These non-regular gaits can make up a significant portion of the daily step count for people, and an improvement to measuring these gaits can drastically improve the performance of the overall pedometer.
Using data collected by 30 participants performing 3 different activities to simulate regular, semi-regular, and irregular gaits, a training and testing strategy was implemented using a sliding window algorithm of pedometer accelerometer data. Data was cut in rows representative of the sliding window, normalized according to the minimum and maximum values of the corresponding sensor-axis combination, and finally collated in specific training and holdout groups for validation purposes.
Nine models were trained to predict a continuous count of steps within a given window, for each fold of our five-fold validation process. These nine models correspond to each gait and sensor combination from the collected data set. Once models are trained, they are evaluated against the holdout validation set to test for both run count accuracy (RCA), a measure of the pedometers detected step to actual step count, and step detection accuracy (SDA), a measure of how well the algorithm can predict the time of an actual step. These are obtained through an additional post-processing step that integrates the predicted steps per window over time in order to find the total count of steps within a given training data set. Additionally, an algorithm estimates the times when predicted steps occur by using the running count of total steps.
Once testing is performed on all nine models, the process is repeated across all five folds to verify model architecture consistency throughout the entire data set. A window size test was implemented to vary the window size of the sliding window algorithm between 1 and 10 seconds to discover the effect of the sliding window size on the convolutional neural network's step count and detection performance. Again, these tests were run across five different folds to ensure an accurate average measure of each model's performance.
By comparing the metrics of RCA and SDA between the machine-learning approach and other algorithms, we see that the method introduced in this thesis performs similarly or better than both a consumer pedometer device, as well as the three classic algorithms of peak detection, thresholding, and autocorrelation. It was found that with a window size of two seconds, this novel approach can detect steps with an overall average RCA of 0.99 and SDA of 0.88, better than any individual classic algorithm.
Lin, Basil, "Machine Learning and Pedometers: An Integration-Based Convolutional Neural Network for Step Counting and Detection" (2020). All Theses. 3462.