What a crowd!
Problem being addressed
The existing approaches cannot handle huge crowd diversity well and thus perform poorly in extreme cases, where the crowd density in different regions of an image is either too low or too high, leading to crowd underestimation or overestimation.
A novel strategy to address the problem of counting in highly varying crowd density images by first classifying the images into either one of the extreme cases (of very low or very high density) or a normal case, and then feeding them to specifically designed patch-makers and crowd regressor for counting. A novel rule-set engine is developed to determine whether the image belongs to an extreme case. For images of extremely high density, a zoom-in strategy is developed to look into more details of the image; while for images of low-density extreme, a zoom-out based regression is employed to avoid overestimate. The four newly created datasets, each from the corresponding crowd counting benchmark, are used for training and testing of different machine learning algorithms to classify an image as normal, high or low-dense extreme case using its patches classification count. These manually verified datasets will facilitate the researchers in analyzing complex crowd diversity, which is at the core of the crowd analysis.
Advantages of this solution
Extensive experimental evaluations demonstrate that the proposed approach outperforms the state-of-the-art methods on four benchmarks under most of the evaluation criteria.
Possible New Application of the Work
Accurate crowd counting is important during all sorts of open air festivals to ensure security and provide better customer service.
In case of mass protests proper crowd counting is important to evaluate the true number of the protestors; quite often these numbers are very approximate and differ significantly based on the source, which makes some of the official reports non-trustworthy. A proper counting tool can provide unbiased results and help better analyse the situation.
Crowd counting tools are essential for the infrastructure of the cities where mass events (e.g. music festivals, food festivals etc.) are very popular. This methodology should be included in the predictive analytics tools to avoid overcrowding and potentially dangerous situations.
Source URL: #############