Intelligent Deep Learning Enabled Crowd Detection and Classification Model in Real Time Surveillance Videos

Abstract

Recently, security surveillance applications exploited the computer vision based detection and tracking approaches to improve the safety and comfort of humans. A major concern in real time surveillance video tracking is the process of identifying the human crowd behavior and classifying them. It finds useful to alert the crow in case of any disasters and unpredicted events. The investigation of human behavior in crowded surveillance videos is an essential and crucial area of research. The recent advances in Artificial Intelligence (AI) and deep learning (DL) models can be employed for determining the crowd behavior analysis in surveillance videos. With this motivation, this article focuses on the design of intelligent deep learning enabled crowd behavior detection and classification (IDL-CBDC) model in real time surveillance videos. The goal of the IDL-CBDC technique is to detect the crowd and classify it into four classes namely marriage, political, school, and college. Primarily, the IDL-CBDC technique performs preprocessing in two levels namely adaptive median filtering (AMF) technique and contrast enhancement (CE) approach. Besides, a deep instance segmentation approach using PSPNet-101 model is used for the segmentation of input video frames into crowds. Moreover, the black widow optimization (BWO) with residual network (ResNet50) model is applied for the crowd detection and classification process. The design of BWO algorithm helps to properly adjust the hyperparameter values such as learning rate, batch size, number of epochs, and number of hidden layers. In order to ensure the improved performance of the IDL-CBDC technique, a set of simulations take place using an own dataset, gathered from public places. Extensive comparative result analysis reported the supremacy of the IDL-CBDC technique over the other techniques.