Hybrid CNN Models in Deep Learning for Smart Crowd Analytics and Real-Time Headcount Prediction

Abstract

Headcount forecasting is a key factor of crowd analytics, headcount prediction is necessary for event management, public safety and urban planning. Traditional counting methods face complex situations, including occlusions, perspective distortions, and dense gatherings. To specify these tasks, deep learning methods, particularly Convolutional Neural Networks (CNNs), have been useful to increase robustness and accuracy in headcount estimation. In this work, we estimate several CNN-based models, including Cross-Modal Transfer Learning (CMTL), Single-column Fully Convolutional Network (SFCN), Context-Aware Network (CANNet), Multi-Column Convolutional Neural Network (MCNN), TransCrowd, and Congested Scene Recognition Network (CSRNet), along with a hybrid MCNN with the CMTL method. Using the dataset of high-density crowd images, models are trained with the best augmentation and preprocessing techniques to guarantee generality. Experimental results disclose that the MCNN with CMTL hybrid attained the highest accuracy, outperforming separate CNN architectures. These results highlight the performance of hybrid CNN representations in developing reliable, scalable, and efficient headcount prediction structures for real-world applications.