Crowd counting method based on feature fusion and attention mechanism

Abstract

Aiming at the problem of background noise interference and occlusion in complex crowded crowd scenes, a crowd counting network FANet based on feature fusion and attention mechanism is proposed. By introducing a feature fusion layer and a crowd region recognition module, FANet can effectively eliminate the influence of background interference and occlusion, thereby improving counting performance. As a supplement to the feature extraction network, the feature fusion layer aims to fuse low-level texture features and high-level features to avoid a large amount of loss of features, thereby enabling the model to have higher multi-scale information perception capabilities and improving training efficiency. The crowd region recognition module generates a corresponding attention weight map for the image through convolution and up-sampling operations, and based on this, achieves the purpose of suppressing background interference. Finally, the evaluation was conducted on two data sets. The experiment showed that the MAE of the proposed method on ShanghaiTech and UCF-QNRF achieved 1.1%,3% and 1.1% improvement respectively.