Human Activity Recognition from an Accelerometer on The Chest: Data Transformation, Feature Selection, and Classification Performance

Abstract

The fourth risk factor for global mortality is lack of physical activity (PA). From the past to present, the relationship between public health and sedentary behavior or physical activity has been an interesting topic for scientists. In the past decade, use of accelerometers for recognizing PA has increased significantly. The aim of this thesis is build a new algorithm to recognize eight different static activities and seven dynamic activities from accelerometer data on the chest based on laboratory data.

To conduct this study, we used laboratory data which was collected from 30 healthy people. In order to extract required information for the analytical part, all activities were recorded in video files. After data collection, all activities were labeled. We used first order differencing to remove the effect of participant’s characteristics. Median of angles and the area under the curve were considered as features and used as predictors in classifiers. We performed 81 different random-forest models to evaluate the effect of sample size and time window size in the accuracy of the model.

We achieved 98.2% accuracy in a random-forest model with 5000 sample size in 6 second time window. We found that there is a positive correlation between time window and sample size with accuracy of the random-forest model.

Also, we performed the Support Vector Machine (SVM) algorithm for same sample size and time window. The accuracy of the SVM model was 95.5%. Both models have reliable performance to recognize the activities in fifteen categories. In the next step, based on sedentary behavior and physical activities definitions, we combined some categories and evaluated the performance of our models in the new categories. As a final result, we achieved 98.9% and 97.6% accuracy in seven different categories. The result of random-forest and SVM models demonstrate our features have provided well-separated data in each category. Future research is required to evaluate the performance of these models on the real-life data.