<p>Video-Based Crowd Analytics Using Deep Learning Methods for Density Estimation</p>

Crowd analytics with video-based analytics is a major research area, and crowds are being monitored automatically for surveillance and public safety. The proposed work proposes a frame-based crowd density estimation model that has the ability to work on both image and video inputs with four different machine learning-based density estimation models. The input videos are extracted as frames, and images and frames are pre-processed into grayscale, normalization, and noise removal. The work proposes a crowd density estimator, namely the Gradient-Texture Crowd Density Estimator, Multi-Scale Crowd Density Estimator, Context-Aware Smooth Crowd Density Estimator, and Global-Context Crowd Density Estimator, to compute frame-wise crowd density maps and density scores in the proposed framework. The density values estimated for each frame are aggregated to analyze the progressive crowd behavior. A comparative evaluation is presented based on the metrics such as Mean Absolute Error, Mean Squared Error, density entropy, prediction variance, runtime, as well as density level and density maps. The experimental results underline that the best overall performance is achieved by the Gradient-Texture Crowd Density Estimator, since this method offers stable density estimation with the least computational overhead. The proposed framework provides an efficient, scalable, and real-time performance solution for crowd density estimation in video analytics and smart surveillance systems.