Crowd counting and localization based on codec structure

Abstract

Crowd counting algorithm began to develop towards the combination of counting and localization. However, the existing methods pay little attention to the localization task, so we propose a crowd counting and localization network based on codec structure, named CALNet. At the end of encoding, the network efficiently fuses different scale information through Adaptive Feature Fusion (AFF) module, to weaken the influence of change of head scale and camera perspective. In the decoding part, we designed the Attention Enhancement Upsampling (AEU) module to obtain high-quality regression map, which uses high-level semantics to guide the model to fuse shallow features. In addition, due to the problem that the existing label maps can not give good consideration to the counting and localization tasks, we proposed a new label map generation method. The new label map not only ensures the simple counting mode, but also has excellent localization performance, and it can be easily migrated to the existing counting methods. The experimental results show that our approach can achieve good results in counting accuracy and localization accuracy.