Suspicious Human Crowd Behaviour Detection - A Transfer Learning Approach

Abstract

Video surveillance systems play an important role in the public security sector. These systems are used to detect a variety of activities such as suspicious human behaviours, analyze crowd behaviours, manage road traffic and track vehicles. Due to the difficulty in manual monitoring of such activities, several research attempts were carried out to automate information extraction from surveillance systems using machine learning and deep learning approaches. The unavailability of large video data sets for processing purposes, is one of the common barriers in this research domain, since the majority of the existing videos are untrimmed, unannotated, and may contain ambiguous data. The purpose of this research is to propose a machine learning based transfer learning approach, to solve the limitations with the datasets specifically for accurate detection of violent crowd behavior via surveillance systems. Four state-of-the-art models were tested with varied configurations and the proposed prototype achieved highest model accuracy of 97%.