Large-Scale Crowd Counting by Hierarchical Context Guided Aggregation Network

Taking advantage of the strong feature learning capabilities of convolutional neural networks (CNNs), recent years have witnessed extensive studies on CNN-based crowd counting methods in different crowd scenes. However, the CNN-based methods still cannot achieve the optimal counting performance because of the influence of scale variation and complex backgrounds in large-scale crowd scenes. To deal with these problems, this article proposes a hierarchical context-guided aggregation network (HCGANet), which can gradually aggregate accurate crowd region information to generate the crowd density map that closely approximates the actual crowd distribution. To be specific, a multiscale context extraction module (MCEM) is used to extract wider contextual information at each level, thus alleviating the excessive fusion of features at different scales. Furthermore, a hierarchical context aggregation module (HCAM) is proposed to facilitate the sufficient fusion of semantic and detailed information at different levels. In addition, we also present an attention-guided module (AGM) to guide the proposed method to focus more accurately on the pixel locations in the crowd regions, providing more attention to the crowd regions. We conduct comprehensive experiments on various datasets to evaluate the counting performance of HCGANet in different crowd scenes. The encouraging results indicate that the proposed HCGANet can perform more accurately crowd counting in different scenes compared with the state-of-the-art methods.