03-11-2012, 11:31 AM
DETECTING HEAD ORIENTATION
DETECTING HEAD.pdf (Size: 138.19 KB / Downloads: 20)
INTRODUCTION
The goal of this project is to train a neural network to classify the orientation of a centered image of
a head as either left, right, up, or down. Head pose gives information about what someone is
paying attention to, and as such is important for social interaction and for mediating learning.
Estimating head pose is an active area of research in computer vision. In child development,
sensitivity to head and eye position matures through a reasonably clear series of milestones. The
earliest of these occurs within about six months, when an infant can decide whether someone is
turned to gaze to their left or to their right, but cannot accurately determine their locus of
attention. In this project, attempt to replicate this level of competence with a simple neural network
that can distinguish between faces turned left, right, up and down (see Figure 1).
Network architecture
The raw images were fed almost directly to the input layer of a neural network. This projected to
a single hidden layer with a small number of neurons. This in turn projected to the output layer,
which had a neuron for each of the four possible orientation classifications.