Region-Aware Multiscale Aggregation Networks for Crowd Counting

Scene complexity and scale variation have long been recognized as common problems in crowd counting. Although deep learning has a good performance in crowd counting, it is still challenging to accurately perform the crowd counting task in scenes with large-scale variations or sheer complexity, and a few objects are often mistaken for people. To address these difficulties, we propose a Region-Aware Multiscale Aggregation Network (RMANet) for crowd counting. On the whole, we design a Fine-Grained Module (FGM). Specifically, FGM consists of an adaptive scale pyramid module (ASPM) and a weight recalibration fusion module (WRFM). ASPM is used to extract contextual features within different regions. WRFM adopts region-aware attention (RAA) to further calibrate the feature weights for different regions. Furthermore, the Multi-scale Aggregation Regression Head Module (MARHM) is introduced to make full use of multi-scale and contextual features extracted from images to generate high-quality crowd density maps and perform accurate count estimation. We implemented extensive experiments on three different challenging datasets, i.e., Shanghai Tech, UCF_CC_50, and UCF-QNRF, and the results demonstrate the effectiveness of the method.