High-Altitude Vertical View Crowd Counting Using Seg-Density Fusion Network

Crowd counting shows promise for practical applications, particularly with drones providing novel perspectives. However, images captured from high-altitude views present detailed surroundings, posing difficulties in recognizing crowd characteristics. In this case, traditional dataset-adapted methods encounter challenges in handling background noise, leading to misidentification. To this end, we propose a novel architecture called Seg-Density Fusion Network. The architecture consists of a semantic segmentation branch that outputs valid-invalid mask maps and a density estimation branch that outputs raw predicted density maps. The outputs of the two branches are then fused on a pixel-by-pixel basis. Experiments use classical MCNN, CSRNet, and ResNet as feature extraction networks for density estimation on the VisDrone-CC benchmark dataset. The results show that the counting error was reduced with minimal degradation to the quality of the density map. This highlights the contribution of masks in mitigating false-positive observations caused by background noise.