SCLNet: Spatial context learning network for congested crowd counting

Abstract

Accurate congested crowd counting is a challenging task, especially in complex crowd scenes. Many existing counting models easily fail in such cases. To solve this problem, we propose a spatial context learning network (SCLNet) for congested crowd counting. SCLNet consists of three parts: feature encoder (FE), spatial context learning decoder (SCLD), and density regression module (DRM). FE firstly processes each input image for feature extracting. Then, the extracted features are fed to SCLD for acquiring spatial context information from different depths of the network. Specially, SCLD consists of three dilated attention modules (DAM). Each DAM applies channel attention mechanism (CA) and spatial attention mechanism (SA) to process extracted features for obtaining spatial context information from the channel and spatial dimensions, respectively. Finally, the features with spatial context information are processed by DRM for crowd density estimation. Experiments are conducted on the ShanghaiTech, UCF_CC_50, UCF-QNRF datasets, and the performance of our method is competitive to the other state-of-the-art methods. Besides, we also evaluate SCLNet on the crowd localization task with the UCF-QNRF dataset, and the results demonstrate the effectiveness of our model.