Crowd Counting Network for Risk Early Warning in Exhibition Scenarios

As a densely populated public place, exhibitions feature spatial layouts with multi-area linkage and instantaneous crowd flow mutations. Thus, developing a crowd flow early warning system adapted to exhibition dynamics is a key focus at the public safety and smart exhibitions to avoid risks like local congestion-induced stampedes. In general, two core challenges in exhibition crowd counting: 1) Key dynamic gathering information is hidden in high frequency components, but no correlation mechanism between frequency components and scene has been established; 2) Instant crowd gatherings cause high-frequency local density mutations, leading to time delays and spatial ambiguity of dynamic signals. To solve these, we propose a novel Crowd Counting Network for Risk Early Warning in Exhibition Scenarios with two core modules: 1) A bidirectional feature filtering module optimizes frequency information through low-frequency suppression to reduce redundancy and high-frequency activation to enhancing dynamic signals. 2) A lightweight self-attention module captures long-range dependencies of high-frequency features via frequency-domain self-attention, enabling accurate identification of feature clusters from gatherings. Validated on datasets like ShanghaiTech and UCF-QNRF, this method provides an integrated solution for exhibition security monitoring, promoting crowd counting technology from general scenarios to industry customization.