Abstract: To solve the problems that the difference in crowd image background and the change in crowd scale caused by perspective effect have a serious impact on the accuracy of crowd counting, a multi-level fusion and attention mechanism based crowd counting algorithm is proposed, which includes two sub networks: scale attention extraction and multi-level fusion. The scale attention extraction network adopts coder-decoder structure, which is responsible for scale extraction to combat the problems of crowd scale change and crowd occlusion in complex crowd scenes; the multi-level fusion network adds a feature fusion operation before each convolution block to fuse the attention map with the input of each convolution block to remove the redundant image information, and then generate a high-quality crowd density map. Compared to other excellent crowd counting algorithms, the MAE(Mean Absolute Error) and MSE(Mean Squared Error) of the proposed algorithm on the ShangHaitech dataset Part _ B are increased by 17% and 25% , respectively, and the MAE on Part _ A is increased by 1. 7% . The MAE is increased by 7% on the UCF_CC_50 dataset. The experimental results show that the proposed algorithm has high accuracy and robustness in dealing with complex crowd scenes.

Key words: crowd counting,  , coder-decoder,  , scale attention,  , feature fusion,  , crowd density map