Fully Convolutional Network for Crowd Size Estimation by Density Map and Counting Regression

Abstract

To handle the customer distribution in the certain areas, crowd counting is necessary for such applications, which is a labor-intensive work for human. Therefore, an automatic crowd counting system is in great demand, but it is still a challenging problem since the human heads and bodies are usually highly overlapping in crowd images. In this paper, a counting-by-regression framework is employed. The human head is modeled as a Guassian distribution. With a crowd density map estimator, the head count can be obtained by integrating over the density map. Most existing approaches only apply density map regression for training a density map estimator, but it is hard to find a suitable training parameters to train a good one; actually, the head count is overestimated easily. To mitigate this problem, counting regression is combined with density map regression. A deeper and lighter fully convolutional network (FCN) is designed to be a crowd density map estimator. The input and output size of the FCN are the same. After training by the proposed method, our model is more competitive comparing with others. The parameter quantity of the model is the lowest, and it needs the least inference time.