Scale Attentive Aggregation Network for Crowd Counting and Localization in Smart City

Abstract

Recent years have witnessed a remarkable proliferation of applications in smart cities. Crowd analysis is a crucial subject, and it incorporates two subtasks in smart city systems, i.e., crowd counting and crowd localization. Nevertheless, the presence of adverse intrinsic factors, i.e., scale variation and background noise severely degrades the performance of counting and localization. Although great efforts have been made on separate research on counting and localization, few works are capable of performing both tasks at the same time. To this aim, the scale attentive aggregation network (SA²Net) is proposed to solve the problems of scale variation and background noise in crowd counting and localization tasks synchronously. Specifically, the SA²Net has two vital modules, namely multiscale feature aggregator (MFA) module and background noise suppressor (BNS) module. The MFA module is designed in a four-pathway structure, and it aggregates the multiscale feature so as to facilitate the correlation between different scales. The BNS module utilizes the contextual information between the input keys matrix and self-attention matrix to suppress the background noise. Furthermore, a global consistency loss combined with the Euclidean loss is utilized to optimize the network in counting and localization tasks. Extensive experimental results prove that the SA²Net outperforms the state-of-the-art competitors both subjectively and objectively.