ARCHIVES

This article discusses a real-time crowd monitoring system developed by deep learners to estimate the crowd density, give automated warnings regarding the occupancy threshold, and create easy-to-use heatmap images to manage public safety. The proposed system uses the Swin Transformer architecture as its key feature extraction backbone, taking advantage of shifted window self-attention to extract both local texture fine-grained and global spatial relationships. Four stages of hierarchical multi-scale features of the transformer are decoded using transposed convolution layers to restore high-resolution Gaussian density maps. A parallel regression head is used to simultaneously predict the total headcount as a scalar. They used the ShanghaiTech Part A and Part B datasets to train and did further fine-tuning using JHU-Crowd++ to enhance cross-domain generalization. A hybrid loss that is a mean absolute error on the density map with a count consistency penalty stabilizes convergence and discourages systematic overcounting. The trained model is applied as a web application in Flask that takes in crowd photographs and displays an annotated heatmap of the JET-colormap that is a heatmap, the approximate number of people, and a live SAFE or DANGER occupancy warning. Experimental performance on ShanghaiTech Part A shows that it is competitive compared to existing convolutional and transformer-based approaches, and the entire deployment pipeline including occupancy alerting is what makes this system stand out compared to algorithmic contributions of the existing literature.