Nested UNet GAN for Enhanced Crowd Counting via Density Map Generation and Accuracy Optimization

The paper proposes Unet Generative Adversarial Networks (GANs) for accurately calculating the number of people in crowd surveillance footage and intelligent transport, where challenges such as high occlusions and cluttered backgrounds often exist. The research focuses on developing a Nested UNet Generative Adversarial Network (Nested UGAN) to perform counting across diverse crowd environments. The nested GAN with UGAN generators and discriminators enhances crowd counting by improving the accuracy of density map generation, leading to more precise differentiation between real and fake crowd images. The findings of the proposed work reveal that the Nested UnetGAN significantly enhances crowd counting accuracy, achieving an MAE of 6.15 and an MSE of 0.19. The proposed method's effectiveness was thoroughly evaluated, demonstrating that it significantly outperforms existing crowd-counting techniques while producing high-quality density maps and effectively handling challenges such as occlusions and varying crowd densities. Additionally, the model's advanced training methods and attentional mechanisms contribute to its robustness and superior performance across ShanghaiTech- A dataset.