Group-split attention network for crowd counting

摘要

Crowd counting is a considerable yet challenging task in intelligent video surveillanceand urban security systems. The performance has been significantly boosted along with thespringing up of the convolutional neural networks (CNNs). However, accurate and efficientcrowd counting in congested scenes remains under-explored due to scale variation and clutteredbackground. To address these problems, we propose a biologically inspired crowd countingmethod named group-split attention network (GSANet). The GSANet consists of three principalmodules, namely GS module, dual-aware attention module, and aggregation module. The GSmodule processes the subfeatures of each group in parallel, and groups the input feature map toreduce the computational cost. The dual-aware attention module synergies the spatial and channeldimensional information to alleviate the estimation error in background regions. The aggregationmodule adopts a learning-based cross-group strategy to aggregate and facilitate the fusionof feature maps along different channel dimensions. Extensive experimental results on fivebenchmark crowd datasets demonstrate that the GSANet achieves superior performances interms of accuracy and efficiency.