Finding Real-Time Crime Detections During Video Surveillance by Live CCTV Streaming Using the Deep Learning Models
Ramesh Chandra Poonia, Kamal Upreti, P. Prabu, K. Rajendra Prasad
Abstract
Nowadays, securing people in public places is an emerging social issue in the research of real-time crime detection (RCD) by video surveillance, in which initial automatic recognition of suspicious objects is considered a prime problem in RCD. Dynamic live CCTV monitoring and finding real-time crime activities by detecting suspicious objects is required to prevent unusual activities in public places. Continuous live CCTV video surveillance of objects and classification of suspicious activities are essential for real-time crime detection. Deep training models have greatly succeeded in image and video classifications. Thus, this paper focuses on the use of trustworthy deep learning models to intelligently classify suspicious objects to detect real-time crimes during live video surveillance by CCTV. In the experimental study, various convolutional neural network (CNN) models are trained using real-time crime and non-crime videos. Three performance parameters, accuracy, loss, and computational time, are estimated for three variants of CNN models for the real-time crime classifications. Three categories of videos, i.e., crime video (CV), non-crime video (NCV), and weapon-crime video (WCV), are used in the training of three deep models, CNN, 3D CNN, and Convolutional Long short-term memory (ConvLSTM). The ConvLSTM scored higher accuracy, lower loss values, and runtime efficiency than CNN and 3D CNN when detecting real-time crimes.