Switching Detection and Density Regression Network for Crowd Counting

Abstract

Counting pedestrians or objects with significantly varying scales and densities has attracted much interest from the research community and yet it remains an open problem. For crowd counting methods, the detection based and the density regression methods have their advantages and disadvantages in terms of scene complexity and head scale. Therefore, using only one of these methods may not be suitable for pedestrian counting with different densities. Given this current issue, we propose a novel crowd counting framework combining both detection-based and density-based methods, which is named SDDRNet (Switching Detection and Density Regression CNN). First, we divided the entire image into several blocks. Each picture block was labeled by a switch classifier according to the crowd density. Then, for low-density image blocks, we used detection-based networks to predict the number of pedestrians, and for high-density image blocks, we used density-regression-based networks to predict the number of pedestrians. Finally, we fuse the results of crowd counting across all of the blocks to get the final output. We evaluated the proposed method on a realistic rail-way station pedestrian image dataset. The experimental results demonstrate that our proposed method outperforms several tested state-of-the-art approaches in overall performance.