Deep Learning-Based Crowd Density Estimation Using Convolutional Neural Networks for Real-Time Analysis

Abstract

Estimating the number of people present in a crowded scene from an image is a challenging computer vision problem, particularly under conditions of severe occlusion, scale variation, and non-uniform crowd distribution. This paper presents a deep learning framework for crowd density estimation that predicts a pixel-wise density map from an input image and derives the total crowd count by integrating over that map. Three convolutional neural network variants are studied and compared: a baseline CSRNet-style dilated-convolution network, a CSRNet variant initialized with VGG16 transfer learning, and a Context-Aware Network (CANNet) that incorporates an attention mechanism to adaptively weight multi-scale contextual features. Ground-truth density maps were generated from point annotations using Gaussian kernels, and the models were trained with a downsampled Euclidean (mean squared error) loss to reconcile the resolution mismatch between the predicted and ground-truth maps. Data augmentation strategies, including horizontal flipping and brightness variation, were used to improve generalization. Experiments conducted on the ShanghaiTech and UCF_CC_50 datasets show that the CSRNet model with VGG16 transfer learning achieves the lowest error, with a Mean Absolute Error (MAE) of 0.38 and a Mean Squared Error (MSE) of 0.69, outperforming the baseline CNN and CANNet configurations. The results demonstrate that transfer learning combined with dilated convolutions is an effective strategy for accurate, scalable crowd density estimation suitable for real-time public safety and surveillance applications.