Modeling of Performance Metrics in Activity Recognition Chains

Type: Semester thesis
Student: Ioannis Tzanos
Advisor: Claudia Villalonga, Dr. Daniel Roggen
Project: OPPORTUNITY, SENSEI

Nowadays, activity recognition systems determine the activity that a user is performing by using a chain of well-known preconfigured sensors, pre-processing functions, features, classifiers and fusion methods. In the future, due to the increase of available sensors in the environment, more than one processing chain could be invoked in order to perform the same activity recognition. The main challenge would be selecting the best components to perform the detection, e.g. the ones that provide the best accuracy in the resulting context or the ones that demand less power consumption to perform the recognition.
In order to allow the automatic creation of recognition chains that fulfill some quality requirement, the elements in the recognition chain should be well characterized, their main properties should be identified and quantified, and the algorithms that allow the selection of the most suitable components should be studied.