Abstract: Crowd count estimation is the linchpin of the crowd management system, which is very important to prevent stampede accident and guide crowd. It has become an increasingly important task and challenging research direction. This paper proposes a data-dependent encoder-decoder architecture with split-attention for crowd counting, called DNeStCount. In order to cope with the challenges of scale variation and perspective distortion of video surveillance, a more dense atrous ratio is applied to the design of the dense atrous spatial pyramid pooling block. In order to improve the accuracy of density map estimation, a learnable and data-dependent upsampling method DUpsampling is applied to the design of the data-dependent feature aggregation. In order to compensate outlier sensitive and untrainable Euclidean loss, Smooth L1 loss is used to the design of loss function. The experiments and analyses on challenging datasets show that DNeStCount is more competitive compared to thoughtful approaches.

Key words: crowd counting, encoder-decoder architecture, split-attention mechanism, dense atrous spatial pyramid pooling； data-dependent upsampling； Smooth L1 loss