Constellation, le dépôt institutionnel de l'Université du Québec à Chicoutimi

Résumé

This paper presents an embedded system that performs activity recognition in the city. Arduino Due boards with infrared, distance and sound sensors are used to collect data in the city and the activity, profile, and group size recognition performance of different machine learning algorithms (RF, SVM, MLP) are compared. The features were extracted based on fixed-size windows around the observations. We show that it is possible to achieve a high accuracy for binary activity recognition with simple features, and we discuss the optimization of different parameters such as the sensors collection frequency, and the storage buffer size. We highlight the challenges of activity recognition using anonymous sensors in the environment, its possible applications and advantages compared to classical smartphone and wearable based approaches, as well as the improvements that will be made in future versions of this system. This work is a first step towards real-time online activity recognition in smart cities, with the long-term goal of monitoring and offering extended assistance for semi-autonomous people.