DLMP-Net: A Dynamic Yet Lightweight Multi-pyramid Network for Crowd Density Estimation

Abstract

The current deep neural networks used for crowd density estimation face two main problems. First, due to different surveillance distance from the camera, densely populated regions are characterized by dramatic scale change, thus using vanilla convolution kernels for feature extraction will inevitably miss discriminative information and reduce the accuracy of crowd density estimation results. Second, popular networks for crowd density estimation still depend on complex encoders with a large number of parameters, and adopt fixed convolutional kernels to extract image features at different spatial positions, resulting in spatial-invariance and computation-heavy. To remedy the above problems, in this paper, we propose a Dynamic yet Lightweight Multi-Pyramid Network (DLMP-Net) for crowd density estimation. The proposed DLMP-Net mainly makes two contributions. First, we design a shuffle-pyramid feature extraction and fusion module (SPFFM), which employs multi-dilated convolution to extract and fuse various scale features. In addition, we add group and channel shuffle operation to reduce the model complexity and improve the efficiency of feature fusion. Second, we introduce a Dynamic Bottleneck Block (DBB), which predicts exclusive kernels pixel by pixel and channel by channel dynamically conditioned on an input, boosting the model performance while decreasing the number of parameters. Experiments are conducted on five datasets: ShanghaiTech dataset, UCF_CC_50 dataset, UCF_QRNF dataset, GCC dataset and NWPU dataset and the ablation studies are performed on ShanghaiTech dataset. The final results show that the proposed DLMP-Net can effectively overcome the problems mentioned above and provides high crowd counting accuracy with smaller model size than state-of-the-art networks.