Scene Adaptive Segmentation for Crowd Counting in Population Heterogeneous Distribution

Abstract

Crowd counting is an important part of crowd analysis and has been widely applied in the field of public safety and commercial management. Although researchers have proposed many crowd counting methods, there is little research on non-uniform population distribution. In this research, a new scene adaptive segmentation network (SASNet) is proposed that can focus on crowd area to estimate accurately crowd density in population heterogeneous distribution. First, an image segmentation module is designed that can adaptive horizontal segment an image according to different density levels, and then obtains a close-up view image and a distant view image. Second, a dual branches network based on convolution neural network (CNN) is exploited that contains a distant view network (DVNet) and a close-up view network (CVNet), so as to extract different scales of image features and then generate density maps by each branch, respectively, so that the crowd counting module has robustness on different scales of target. Finally, a comparative experiment on three well-known crowd counting datasets shows that SASNet achieved stabilized performance and robustness in population heterogeneous distribution.