Generation of High Quality Density Map Using USkipGAN

In intelligence video surveillance, crowd analysis tasks like counting are gaining much importance. It is applied in urban planning, traffic management, and avoiding untoward situations due to overcrowding. The main challenge lies in considering varying crowd sizes, distances, and features while precisely counting and building a high caliber density map. Considering the above challenges in the crowd, this paper presented UskipGAN architecture comprising Unet skip connection, Unet generator, and discriminator to construct a good quality density map. The Unet skip connection fabricated as an encoder-decoder structure executes semantic segmentation by including global and local spatial features. The unet generator concatenates input from the Unet skip connection to its varying kernel size to detect finer and more diverse details of the people. The resulting density maps are refined in the discriminator by deploying a binary entropy loss function. The observation is evaluated on the datasets of ShanghaiTech A, B, and UCFF_CC_50, having varied scenes and densities deploying mean square error (MSE) and mean absolute error (MAE) as metrics. The preliminary test proves the recommended model's effectiveness on the MSE metric on the above three datasets, and it outperforms on UCFF_CC_50 datasets.