Single Image Crowd Counting Based On Deep Feature Fusion

The stampede event is one of the most dangerous accidents in daily life.It brings a great threat to the survival of human beings.It is essential for modern urban construction to suppress the occurrence of stampede accidents effectively.At present,it is possible to effectively analyze abnormal events in the scene and provide early warning of dangers by combining crowd density estimation(or crowd counting)technology with crowd behavior analysis.This is one of the most effective measures to prevent stampede events.However,the crowd counting task is easily interfered by factors such as perspective,crowding,non-uniform and occlusion in the image,which results in large changes in the scale and appearance of the human head in the image.These problems makes accurate crowd counting a challenging task.To solve these problems,this paper proposes a crowd counting algorithm based on the deep feature fusion models.It is achieved by mining the correlation of hierarchical information and the multi-scale and multi-semantic features in complex scenarios.The main works of the paper are as follows:1)Investigating the importance of scale information for crowd counting,this paper proposes a hierarchical multi-scale fusion module to effectively fuse different scale information in the network.Which achieve more accurate crowd counting.2)Considering that the crowd counting is to estimate the specific target in an image and this is easy to be interfered by the surrounding environment,this paper proposes to use an attention mechanism module to weight the feature map in a spatial position,thus effectively limiting the interference of the surrounding environment.3)For crowd counting,the current evaluation criteria,MAE and MSE,only focus on the total number of people,ignoring the rationality of the density map.The introduction of PSNR and SSIM is to measure the correlation between the estimated density map and the label,which is contrary to the inaccuracy of the label data.In this paper,we introduce a patch-based metric to measure the rationality of density map.Finally,the experimental results show that the proposed model can achieves competitive performance on several public datasets compared to the other algorithms and this method also showed that our model has a strong generalization ability.