Direction-aware attention aggregation for single-stage hazy-weather crowd counting.

Crowd counting in adverse hazy weather is inevitable and significant to the scene understanding in real-world application. For the crucial and challenging hazy-weather crowd counting issue, this paper proposes a single-stage hazy-weather crowd counting method based on direction-aware attention aggregation. The method develops a Directional Context Attention (DCA) block to realize embedding the positional information to the channel attention aggregation and to capture the long-range dependencies about the crowd structure, both of which guide the counting method readily locate the crowd of interest. Due to the lack of the hazy-weather crowd counting benchmark datasets, we also generate the hazy-weather crowd counting datasets to evaluate the proposed method. Experimental results and the discussions demonstrate our method is effective greatly on the hazy-weather crowd counting task, especially the design of the core DCA block. In this research work, both the proposed method and the generated benchmark datasets in hazy-weather scene facilitate the development of the crowd counting in the adverse weather conditions and the universal intelligent surveillance technologies. The generated hazy-weather crowd counting benchmarks datasets (Hazy-ShanghaiTechRGBD and Hazy-JHU) in this paper would be released in https://github.com/312524/Hazy-CC after peer-review process of this work to facilitate further research.