The objective of this 24 person-months project is to develop a state-of-the-art multi-camera people detection algorithm combining deep learning and metric learning in a consistent framework.
While there has been an intense research activity in both monocular people detection with deep-learning techniques, and multiple-camera detection using more classical computer-vision methods, there is currently no standard algorithm for combining both approaches in a common framework.
We propose to combine multiple fine-tuned neural networks trained for monocular people detection into a multi-camera setup by using a method derived from our large-scale metric learning algorithm: Instead of aiming at identifying individuals over different views in time and space, the metric learning will here serve the purpose of checking if the different views are consistent and correspond to the presence of an individual at a certain geometrical position.