Hybrid High-order Attention-aware Network For Crowd Counting

In the field of crowd counting,the deep features based on the convolutional neural network have achieved excellent counting performance.Almost all the latest crowd counting algorithms use a density-map-based regression method to achieve crowd counting.It retains the spatial distribution information of the crowd and intuitively provides crowd density in different areas.In a real crowd scene,to count accurately,there are lots of challenges,including scale variations caused by the perspective effect,uneven crowd distribution,and the complex interactions between the targets in the scene.However,existing methods either cannot effectively extract multi-scale crowd features,or only rely on first-order attention mechanisms(such as 2D position-wise attention),and completely ignore the higher-order statistics in crowded scenes.In order to solve the above problems,this paper first proposes an encoding network,which uses dilated convolution to significantly expand the receptive field of the network without introducing extra calculations,and extracts deep crowd features in a larger context.Then introduces the decoding network,which uses dense connections with the encoding network,and uses distributed supervisions to better deal with the problem of scale variations.In the encoding and decoding stages,different attentions are introduced in this paper,namely higher-order attention module and adaptive compensation loss function.On the one hand,the higher-order attention module captures higher-order interactions about scene targets in the form of kernel functions,and finally produces a 3D attention map,which focuses on key areas,filters useful information,and extracts discriminative features around the people.On the other hand,under the distributed supervisions,the adaptive compensation loss function utilizes prior knowledge from higher layers to guide the prediction of lower-level density maps and adaptively focus on different regions.Therefore,these components establish a hybrid attention-aware network for crowd counting.Extensive experimental results on three challenging benchmarks show that the hybrid attention-aware network proposed in this paper has achieved the state-of-the-art performance.