An attention mechanism-based multi-scale network crowd density estimation algorithm

It is becoming more and more important to calculate the people number in terms of the requirement for the safety management, because that the crowd gathering scenes are common whether or not it is daily urban traffic or some special gatherings. Calculating the people number in high-density crowd is a very difficult challenge due to the diversity of ways people appear in crowded scenes. This paper proposes a multi-branch network which combines the dilated convolution and attention mechanism. By combining dilated convolution, the context information of different scales of the crowd image are extracted. The attention mechanism is introduced to make the network pay more attention to the position of the head of the crowd and suppress the background noise, so as to obtain a higher quality density map. Then add all the pixels in the density map to get the total people number. Through a large number of experiments, this network can better provide effective crowd density estimation features and improve the dissimilarity of density map distribution, which has stronger robustness.