Offre de stage Recherche - ENS - Oct 2019

Deep learning methods for the sensor-based activity recognition tasks

Deep learning can largely relieve the effort on designing features and can learn much more high-level and meaningful features by training an end-to-end neural network. Ambient sensors are used to capture the interaction between humans and the environment. They are embedded in the user’s smart environment.

The project deals with the evaluation of machine learning methods, based on a dataset from a smart home (co-working space), about activities observed over 20 working days. The dataset is proposed by Orange and INRIA (paper “A Dataset of Routine Daily Activities in an Instrumented Home” by UCAmI 2017). The idea is to evaluate different deep learning algorithms, for different purposes:

• Recognition of atomic activities;
• Learning scenarios;
• Sensor clustering for given activities / scenarios;
• Learning similar sequences

The learning process could be completed with more or less predefined expert knowledge (ex.: labeling, location of sensors).

Depending on the candidat's motivation and results, the work could be followed with several additional tasks:

• Identification of abnormal situations;
• Further activities prediction of further values for the sensors;
• Adaptation to new configurations: to understand the world from the data of a set of sensors, for which we learned activities beforehand (adding / removing sensors, for instance in the case of defective sensors or adding new connected objects).

Mots-clés : sensor-based HAR, smart environment, pervasive computing

Scikit-learn, Python librairy for Machine Learning, INRIA

veronique dot deslandres at liris dot cnrs dot fr