A Novel cascaded deep architecture with weak-supervision for video crowd counting and density estimation.

Autor: Tripathy, Santosh Kumar, Srivastava, Subodh, Bajaj, Divij, Srivastava, Rajeev
Předmět:
Zdroj: Soft Computing - A Fusion of Foundations, Methodologies & Applications; Jul2024, Vol. 28 Issue 13/14, p8319-8335, 17p
Abstrakt: Video-based crowd counting is an essential surveillance tool that plays a crucial role in mitigating crowd catastrophes by facilitating the development and implementation of efficient crowd management methods. The deep learning approaches using density map-based regression consider local crowd distribution but are erroneous for point-level annotation of human heads. The weakly supervised approach overcomes such an issue by mapping global crowd attributes onto ground-truth counts. Also, video-based density map regression approaches don't handle human shape variation and background effects. Hence, this research suggests a unique cascade of two deep structures: a local density map regressor and a global crowd count regressor with weakly supervised learning. The former model can effectively deal with human shape variation, minimise background effects, consider local crowd distribution, and provide crowd density maps. In contrast, the latter adopts a weakly supervised learning mechanism and provides scene-level crowd counting by considering global attributes of density maps. The trials were conducted using three datasets, namely Venice, Mall, and UCSD, yielding promising and improved outcomes. The codes can be available at https://github.com/santosh1448/LDR_GCCR_Weakly_Supervised. [ABSTRACT FROM AUTHOR]
Databáze: Complementary Index
Externí odkaz: