People Detection in a Depth Sensor Network via Multi-View CNNs trained on Synthetic Data

In this work an approach for wide-area indoor people detection with a network of depth sensors is presented. We propose an end-to-end multi-view deep learning architecture which takes three foreground segmented overlapping depth images as input and predicts the marginal probability distribution of people present in the scene. In contrast to classical data-driven approaches our method does not make use of any real image data for training but uses a randomized generative scene model to generate synthetic depth images which are used to train our proposed deep learning architecture. The evaluation shows promising results on a publicly available data set.