Human crowd behaviour analysis based on video segmentation and classification using expectation-maximization with deep learning architectures

In recent years, the demand for automatic crowd behavior analysis has surged, driven by the need to ensure public safety and minimize casualties during events of public and religious significance. However, effectively analyzing the nonlinearities present in real-world crowd images and videos remains a challenge. To address this, research proposes a novel approach leveraging deep learning (DL) architectures for the segmentation and classification of human crowd behavior. Our method begins by collecting input from surveillance videos capturing crowd activity, which is then processed to remove noise and extract the crowd scene. Subsequently, we employ an expectation-maximization-based ZFNet architecture for accurate video segmentation. The segmented video is then classified using transfer exponential Conjugate Gradient Neural Networks, enhancing the precision of crowd behavior characterization. Our method has been proven effective in experimental analysis on many human crowd datasets, with significant results of average mean precision (MAP) of 59%, the mean square error (MSE) of 61%, accuracy in the training of 95%, validation precision of 95%, and selectivity of 88%. The potential of DL-based methods to advance crowd behavior analysis for improved privacy and security is highlighted by this study.