This project will focus only in tracking the head in color images. The 2D head location along with the camera parameters will be used in computing the 3D position of the head. The module that tracks the location of the head in an image has to be autonomous and be robust enough in the environment where the cameras are installed, namely an office. Some requirements of the final application (activity recognition)include robustness respect to the background. While the background in an office is mostly fixed, the person will be moving in the scene and the tracker window will effectively deal with a changing background. The algorithm has to tolerate out of plane rotations (tilt and jaw) translation and scale changes.
The approach proposed by Birchfield meets the requirements and is easy to implement. The main assumption made by using Birchfield approach is that the head can be modeled as an ellipse. The HAR system only needs the trajectory of the head's centroid. Approximating the head by an ellipse is sufficient for finding its centroid. One last assumption is that illumination changes are not very strong. The environments for which the HAR system will be built are buildings indoors, such as a school, hospital, airport or an office. These type of environment have their own source of light and is normally fixed. The only way illumination becomes a problem is when the subject being tracked is illuminated with one source and then moves towards another source. Since inside buildings lighting is highly diffuse then illumination changes won't become a problem.
The goal of this project is to implement the algorithm described in [1] and to test it with image streams from indoor environments.