RGB-T Crowd Counting from Drone: A Benchmark and MMCCN Network

Tao Peng, Qing Li, Pengfei Zhu; Proceedings of the Asian Conference on Computer Vision (ACCV), 2020

Crowd counting aims to identify the number of objects and plays an important role in intelligent transportation, city management and Security monitoring. The task of crowd counting is much challenging because of scale variations, illumination changes, occlusions and poor imaging conditions, especially in the nighttime and haze conditions. In this paper, we present a drone based RGB-Thermal crowd counting dataset (DroneRGBT) that consists of 3600 pairs of images and covers different attributes, including height, illumination, density. To exploit the complementary information in both visible and thermal infrared modalities, we propose a multi-modal crowd counting network (MMCCN) with a multi-scale feature learning module, a modal alignment module and an adaptive fusion module. Experiments on DroneRGBT demonstrate the effectiveness of the proposed approach. We also provide a new thought to the field of RGB-T translation for crowd counting.