Unsupervised Methods for Camera Pose Estimation and People Counting in Crowded Scenes

Most visual crowd counting methods rely on training with labeled data to learn a mapping between features in the image and the number of people in the scene. However, the exact nature of this mapping may change as a function of different scene and viewing conditions, limiting the ability of such supervised systems to generalize to novel conditions, and thus preventing broad deployment. Here I propose an alternative, unsupervised strategy anchored on a 3D simulation that automatically learns how groups of people appear in the image and adapts to the signal processing parameters of the current viewing scenario. To implement this 3D strategy, knowledge of the camera parameters is required. Most methods for automatic camera calibration make assumptions about regularities in scene structure or motion patterns, which do not always apply. I propose a novel motion based approach for recovering camera tilt that does not require tracking. Having an automatic camera calibration method allows for the implementation of an accurate crowd counting algorithm that reasons in 3D. The system is evaluated on various datasets and compared against state-of-art methods.