HMMs and CRFs for continuous human activity recognition
The main mechanism that humans have for learning is the demonstration and imitation of activities. For instance, we learn how to talk listening to our parents when we are a baby, and trying to repeat everything when we are a bit older. But, can we make that a robot learns in the same way? Nowadays a technique called programming by demonstration (PbD) has gained popularity. Its main concept is that a teacher shows a robot how to do something but, instead of writing code, repeating the same action. Then the robot must extract the movements performed by the teacher, and later it must be able to understand the meaning of those movements and generalize to unseen cases.
In this work we develop two methods to recognize and understand actions performed by that teacher. We have to be able to segment the sequence of raw data obtained by the robot in order to get a sequence of actions. The first method uses a Hidden Markov Model (HMM), which is a probabilistic framework, for recognition purposes, while an algorithm called level building is used for segmenting the sequence of raw information. The second possibility we consider is the use of another stochastic framework called Conditional Random Field (CRF), which is able to segment and recognize actions at the same time.