Fusion of Octave and Dilated Convolutional Networks for Enhanced Dense Crowd Counting

Accurate crowd counting is crucial for numerous applications ranging from public safety to urban planning. However, existing methods often struggle to adequately capture both fine-grained frequency-domain information and broader semantic context cues, limiting their performance in accurately estimating crowd density. To overcome these problems, we propose a novel method that harnesses the strengths of octave convolutions and dilated convolutions. Octave convolutions effectively decompose features into high-frequency and low-frequency components, focusing on detailed textures of the crowd distribution. Meanwhile, dilated convolutions expand the receptive field without increasing number of parameters, and enable our model to capture rich semantic context for crowd estimation. The experimental results demonstrate that our approach can achieve precise crowd counting by jointly learning frequency-based details and semantic information. Extensive experiments show that the proposed method achieves the competitive performance on ShanghaiTech datasets for crowd counting.