Leveraging deep learning with symbolic sequences for robust head poses estimation

Head pose estimation is a challenging topic in computer vision with a large area of applications. There are a lot of methods which have been presented in the literature to undertake pose estimation so far. Even though the efficiency of these methods is acceptable, the sensitivity to external conditions is still being a big challenge. In this paper, we come up with a new model to overcome the problem of head poses estimation. First, the face images are converted into one-dimensional vectors as a time series using the Peano-Hilbert space-filling curve. Then, we convert these numerical series into symbolic sequences with adequate dimensionality reduction approaches. These sequences are then used as input of an encode-decoder neural network to learn and generate labels of the faces orientations. We have evaluated our model on several databases, and the experimental results have shown that the proposed method is very competitive compared to other well-known approaches.