Human Head Detection and Tracking

In this report, I present a novel algorithm to detect and track the head regions in a normal human walking sequence. The algorithm also classifies each head orientation for the detected head regions. The main assumption for the algorithm is: Head is always located at an extreme position of the body image and there is a noticeable difference between the head and shoulders in the input images. There are several systems used to detect the head: Ismail Haritaoglu et al. used the geometry features to segment the head region [3], Saad Ahmed Sirohey used an ellipse to represent the head region in an edge map [2], Sangho Park et al. used a partial ellipse to fit the head contour in the silhouette image[1]. Combined with these models, I first apply a bounding box for the head region, and then use an extended ellipse model to match the head region and perform the recursive convex hull algorithm to segment the head region. Kalman Filter is applied on the ellipse representation to track the head. Furthermore, I implement and improve the algorithm of Sangho Park [1] to classify the head orientation.