面向跨域人群计数的头部感知密度适应网络

The crowd counting based on the domain adaptive method is an effective unsupervised learning strategy, which does not rely on labeled samples. However, the existing methods easily cause information loss in the head region or over counting errors in the background region. A head-aware density adaptive network (HADAN) is proposed for cross-domain crowd counting to solve these problems. In style transform part, the ground-truth of the source domain dataset is firstly used to generate the mask of the head area and background area, and then a head-aware cycle loss based on the mask is designed to prevent the confusion between the head area and the background area during the style transform process. Simultaneously, the density adaptation part further maps the features of both the source domain and the target domain to the same latent space using the discriminator, enhancing the consistency of the density map distribution. Proposed network trains style transfer and density adaptation parts in an end-to-end way, and they learn iteratively and benefit from each other. The experimental results on the synthetic data set GCC and three real-world datasets show that the MAE value of the algorithm is reduced by 9% on average, and the MSE value is reduced by 7%. Proposed method achieves robust cross-domain crowd counting in unlabeled target scenes.