Crowd Density Estimation Based on Multi-level Attention Maps

Focusing on the crowd counting task in crowded scene, a crowd counting model with U-Net architecture is designed based on multi-level attention maps, which is composed of encoder, decoder and branch network. In the encoder network, The VGG16 convolution block and channel attention module are used for extracting the image features. In the decoder network, 5 Decoder Modules are stacked to generate 5 crowd density maps with different scales, 5 attention maps with different scales are used to extract the key information of head position and overcome the interference of complex background. The branch network is designed for generating above mentioned attention maps. In optimizing the model, a multi-level supervised learning mechanism is introduced by using crowd density maps of 5 different scales and attention maps of 5 different scales. The model is trained and evaluated on the crowd counting dataset Shanghai Tech. A series of ablation experiments are carried out on the dataset Shanghai Tech PartA to validate the effectiveness of the key components of the model. In addition to Mean Absolute Error(MAE) and Mean Squared Error(MSE), Mean Relative Error(MRE) and Structural Similarity(SSIM) are also adopted in the model evaluation. The experimental results demonstrate that the proposed model can effectively improve the accuracy and robustness of crowd counting compared with some typical crowd counting models.