Descripción detallada

Activity Recognition (AR) is the process that enables our smartphones to know how much we have walked this week or if we went to work by bike yesterday. By analyzing the data produced by sensors, we can identify the activity performed by an user in a given time. Activity Recognition lies at the heart of many context-aware applications in domains such as ambient intelligence, healthcare and wellness, sports analysis, surveillance, personalized advertising and others.
The main goal of Activity Recognition is to automatically identify what the user is doing from sensor data captured over a period of time. Although the AR problem can be framed as a traditional machine learning problem, it has some characteristics that make it a different problem, for example:
• The temporal nature of sensor data does not fit the independent and identically distributed hypothesis of supervised learning problems. Instead, sequential relationships in the data must be considered to improve performance.
• Different durations of activities give rise to imbalanced dataset and natural personal changes give rise to distribution changes over time.
• Differences in the way activities are done from person to person introduce challenges to feature engineering and personalization. As a simple example, one person might walk as fast as another person jogs.
• Complex activities that are a composition of many other activities, the “null activity” (time where no activity of interest is performed) and new activities not considered in the training model introduce challenges for model training and adaptation.
In this course we study the activity recognition chain (data acquisition, preprocessing, segmentation, feature extraction, classification). We study techniques to overcome specific challenges of activity recognition at each step including sensor selection, windowing techniques, feature engineering, feature learning and training models. Many of these techniques can be applied to the analysis of other sequential data (sensor data, stock market data, voice recognition).
We discuss system design considerations such as user interface, handling streaming data and network considerations. At the end, we also discuss some ethical implications of human activity recognition.
The experience of the professors is on the healthcare domain, in applications like elder care and nurse care activity recognition. Examples from this domain will be studied throughout the course.