Denstity Level Aware Network for Crowd Counting

Crowd counting has wide applications in video surveillance and public safety, while it remains an extremely challenging task due to large scale variation and diverse crowd distributions. In this paper, we present a novel method called Density Level Aware Network (DLA-Net) to improve the density map estimation in varying density scenes. Specifically, we divide the input into multiple regions according to their density levels and handle the regions independently. Dense regions (with small scale heads) require higher resolution features from shallow layers, while sparse regions (with large heads) need deep features with broader receptive filed. Based on this requirement, we propose to predict multiple density maps focusing on regions of varying density levels correspondingly. Inspired by the U-Net architecture, our density map estimators borrow features of shallow layers to improve the estimation of dense regions. Moreover, we design a Density Level Aware Loss (DLA-Loss) to better supervise those density maps in different regions. We conduct extensive experiments on three crowd counting datasets (ShanghaiTech, UCF-CC-50 and UCF-QNRF) to validate the effectiveness of the proposed method. The results demonstrate that our DLA-Net achieves the best performance compared with other state-of-the-art approaches.