Crowd Density Estimation Using Enhanced Multi-Column Convolutional Neural Network and Adaptive Collation

Accurate crowd density estimation is essential for public safety operations, urban and transportation design and multiple intelligent systems. The research presents an improved Multi-Column Convolutional Neural Network (MC-CNN) structure to predict crowd density. The model uses MobileNet transfer learning capabilities and implements three parallel convolutional columns with different kernel sizes to extract spatial information at different receptive field levels. The design enhances the ability to extract and combine multi-scale features, resulting in better spatial crowd density representation. The model incorporates the output of its columns to improve its understanding of crowd movement patterns. During training, a dynamic collation and zero-padding strategy is introduced to handle variable-sized input images. This approach’s uniform batch shape improves training stability and allows flexible real-world deployment with inconsistent input resolutions. The experimental results show that the proposed architecture performs better, as revealed by its low MAE scores of 12.93 in Part A and 28.56 in Part B of the ShanghaiTech data, outperforming TransCrowd, CSRNet, and SANet. The research provides a functional solution that accurately measures crowd density in different scenarios while overcoming the existing approaches’ restrictions. It fills the academic research gaps in multi-scale feature extraction, efficient data collation and offers new solutions for crowd analysis.