Please login first

Complex activity recognition using wearable airborne particulate matter and motion sensor data
Rok Novak * 1, 2 , David Kocman 1 , Johanna Amalia Robinson 1, 2 , Tjaša Kanduč 1 , Milena Horvat 1, 2 , Denis Sarigiannis 3, 4, 5
1  Department of Environmental Sciences, Jožef Stefan Institute, Ljubljana, Slovenia
2  Jožef Stefan International Postgraduate School, Ljubljana, Slovenia
3  Environmental Engineering Laboratory, Department of Chemical Engineering, Aristotle University of Thessaloniki, Thessaloniki, Greece
4  HERACLES Research Centre on the Exposome and Health, Center for Interdisciplinary Research and Innovation, Thessaloniki, Greece
5  University School of Advanced Study IUSS, Pavia, Italy


Knowing what individuals are doing when they are exposed to elevated levels of pollution is crucial to implement plans to reduce possible harm. Merging new sensing technologies with machine learning methods can be used as a tool to recognize complex activities. In this work, a novel approach of providing a wearable airborne particulate matter and ambient sensor in combination with a motion tracking wrist device to 97 individuals involved in the ICARUS H2020 project in Ljubljana, Slovenia, was used. They were instructed to wear the devices for 7 days, while they manually recorded their hourly activities. The compiled dataset was cleaned and separated into a training and testing data set (each containing unique participants). As the activity data was in hourly intervals and sensor data in minute values, two approaches were used: a) transforming each hourly activity to 60-minute iterations and b) averaging sensor minute data to hourly values. These data sets were used in three different models, based on three classification algorithms: k-Nearest Neighbors (IBk), decision tree (J48) and random forest (RandomForest). The results of the models for hourly values showed an accuracy of 31.0%, 28.6% and 35.7% for IBk, J48 and RandomForest, respectively, and for minute values 23.1%, 22.0% and 23.0% for IBk, J48 and RandomForest, respectively. As expected, most misclassified instances were observed for activities with vague definitions, such as resting and playing. Low accuracy can also be explained with the differences in time scales. The accuracy could be improved by more clearly defining the activities and collecting minute value data. To this end, this research provides a crucial first step in determining the possibilities of combining information coming from various new sensing technologies for complex activity recognition.

Keywords: pattern recognition, particulate matter, machine learning