HRANet: Hierarchical region-aware network for crowd counting

Abstract

Aiming to tackle the most intractable problems of scale variation and complex backgrounds in crowd counting, we present an innovative framework called Hierarchical Region-Aware Network (HRANet) for crowd counting in this paper, which can better focus on crowd regions to accurately predict crowd density. In our implementation, first, we design a Region-Aware Module (RAM) to capture the internal differences within different regions of the feature map, thus adaptively extracting contextual features within different regions. Furthermore, we propose a Region Recalibration Module (RRM) which adopts a novel region-aware attention mechanism (RAAM) to further recalibrate the feature weights of different regions. By the integration of the above two modules, the influence of background regions can be effectively suppressed. Besides, considering the local correlations within different regions of the crowd density map, a Region Awareness Loss (RAL) is designed to reduce false identification while producing the locally consistent density map. Extensive experiments on five challenging datasets demonstrate that the proposed method significantly outperforms existing methods in terms of counting accuracy and quality of the generated density map. In addition, a series of specific experiments in crowd gathering scenes indicate that our method can be practically applied to crowd localization.

Neural Network-Based Crowd Counting Systems: State of the Art, Challenges, and Perspectives

“…Xie and Gu et al [54] proposed Hierarchical Region Aware net can more effectively concentrate on crowded areas to predict crowd density. The Region Aware Module in HRANet allows adaptive extraction of contextual information inside various regions.…”

“…Crowd counting using weakly supervised learning via CNN typically cannot demonstrate good performance because CNN is not adequate for modelling the global context and the interactions between image patches. To model the overall environment and teach contrast characteristics, the weakly supervised model via Transformer was proposed in [33,54]. The network's parameter number is rather huge, and the transformer directly divides the crowd photos into a collection of tokens, which may not be the best option given that each pedestrian is a separate individual.…”

“…The network's parameter number is rather huge, and the transformer directly divides the crowd photos into a collection of tokens, which may not be the best option given that each pedestrian is a separate individual. Transformer [54] can process input sequences in parallel, allowing it to handle longer sequences more efficiently. The self-attention mechanism of the Transformer allows it to weigh the importance of different parts of the input sequence when generating outputs.…”

“…4-6. [50] 194.1 297.8 Switch CNN [67] 318.1 439.2 LR-CNN [13] 325.6 369.4 CAT-CNN [40] 235.5 324.8 CCTrans [34] 245.0 343.6 JCTNet [54] 222.9 306.5 LibraNet [62] 181.2 262.2 MLPCNN [37] 238.63 317.28 HRANet [56] 160.9 235.8 MANet [59] 240.8 311.5 AUCNN [58] 231.8 312.4 AMSNet [42] 236.5 319.2 ASANet [51] 185.5 268.3 [65] 136.26 240.83 DEAL [50] 100.1 166.5 JCTNet [37] 90 161 TransCrowd [33] 97 168 CCTrans [34] 92 158 LibraNet [62] 88.1 143.7 MLPCNN [37] 103.61 168.9 HRANet [56] 84.6 146.2 EDENet [38] 86.6 158.5 JCCL+R [53] 100 169 AUCNN [58] 112.3 195.6 AMSNet [42] 86.5 167.2 WSITrans [55] 86.5 140. From Tables II-IV and its corresponding graph Figs.…”