Pedestrian head detection using automatic scale selection for feature detection and statistical edge curvature analysis

In this report we focus on pedestrian head detection and tracking in video sequences. The task is not trivial in real and complex scenarios where the deformation induced by the perspective field requires a multi-scale analy- sis. Multi-scale shape models for the human head are considered to identify the correct size of the region of interest. Anisotropic diffusion is used as a pre-processing step and edge detection is performed using an automatic scale selection process. A non parametric statistical description is given for the edge curvature and detection is performed by means of goodness-of-fit tests. The head detector is used as a validation tool in a correlation-based tracker. The local maxima of the correlation matrix are analyzed. Tracking is performed associating the displacement vector of the target with that local maximum which maximizes the goodness-of-fit with the distribution of the edge curvature of the head.