PROJECT TITLE :

Counting People by Estimating People Flows

ABSTRACT:

Deep neural networks are used in contemporary techniques for counting the number of people in crowded scenes. These networks use individual images to estimate the population density. Because of this, only a very small percentage of video sequences take advantage of temporal consistency, and the ones that do so only impose minimal smoothness constraints across successive frames. In this paper, we argue that direct regression of people densities should be avoided in favor of estimating people flows across image locations from one image to the next and then deducing the people densities from these flows. Because of this, we are able to impose significantly stricter constraints, which encode the maintenance of the existing number of people. As a consequence of this, the performance is significantly increased without the need for an architecture that is significantly more complicated. In addition to this, it enables us to make use of the correlation that exists between the flow of people and the flow of optical information in order to further improve the results. We also show that it is possible to train a deep crowd counting model in an active learning setting with much fewer annotations by leveraging people conservation constraints in both a spatial and temporal manner. This makes it possible to train the model with fewer people. This results in a significant reduction in the cost of annotation while still leading to performance that is comparable to the case where full supervision is applied.