Improving Temporal Consistency in Aerial Based Crowd Monitoring Using Bayes Filters

In order to monitor mass events, crowd managers continuously require reliable measurements of the crowd count. For this purpose, a variety of deep learning algorithms has been developed. Most of these so-called crowd counting algorithms return good results for still imagery but return oscillating crowd counts for video data. This is because, most crowd counting algorithms evaluate video data frame by frame and ignore the temporal relation between adjacent frames. In this paper, a variety of Bayesian filters is presented that successfully smooth the oscillating counts which in turn can lead crowd managers to trust the system more. The proposed filters work on top of the crowd counting algorithms' estimates. Thus, they can be easily used with any existing crowd counting algorithm that outputs a density map for a given input image.