CCPTNet: A Crowd Counting Model Based on Point Prediction with Transformers

Abstract

In the process of rapid development of urbanization, the role of people flow monitoring technology in ensuring public safety, optimizing urban management and helping to prevent and control epidemic situation has become increasingly prominent. However, when facing the congested environment of high-density crowd, such as subway stations and large-scale events, traditional methods encounter major difficulties in automatic counting, especially due to target overlap and shape distortion caused by dense crowds and inaccurate counting caused by the changing lighting conditions and the diversity of observation angles. Therefore, this study proposes a novel network model - CCPTNet. This model is based on the point prediction model, which not only improves the counting performance and positioning accuracy, but also makes full use of the pyramid vision transformer trunk to enhance the capture of global data features and effectively reduce the risk of over-fitting of the model. At the same time, in order to reduce the light change and the interference to the crowd count from different perspectives, the author uses the conventional random cropping and horizontal flip, and introduces the methods of random rotation and color jitter to enhance the data. In addition, the function of prediction and counting is realized by point coordinate regression and proposal classification through two parallel branches. The experimental results on the Shanghai science and technology data set show that the CCPTNet model has a good accuracy performance.