PROJECT TITLE :

Locate, Size and Count Accurately Resolving People in Dense Crowds via Detection

ABSTRACT:

We present a detection method for dense crowd counting that replaces the widely used density regression paradigm. Rather of detecting every person, typical counting techniques forecast crowd density for a picture. In general, these regression approaches fail to accurately localize people for most applications other than counting. As a result, we use an architecture that locates each individual in the crowd, sizes the spotted heads using a bounding box, and then counts them. There are certain particular obstacles in creating such a detection system when compared to standard object or face detectors. Some of these are direct results of the enormous variability in dense crowds, as well as the requirement to forecast boxes in a consistent manner. We address these challenges and create the LSC-CNN model, which can dependably detect people's heads in crowds ranging from sparse to packed. To better resolve people and produce refined predictions at several resolutions, LSC-CNN uses a multi-column architecture with top-down feature modulation. Surprisingly, the suggested training regime only requires point head annotation, but it can estimate head size estimates. We show that LSC-CNN not only beats previous density regressors in terms of localisation, but also in terms of counting. Our approach's code can be seen at https://github.com/val-iisc/lsc-cnn.