A DEEP CONVNET FOR VIOLENCE DETECTION USING DYNAMIC MOTION IMAGES

Abstract

Over the last decade, there has been a surge in developing automated intelligent video analysis systems that can monitor human activities in the public environment and recognize abnormalities like violent and suspicious events. Violence Detection is an emerging topic in monitoring human activities. Deploying such violence detection automated systems at highways, shopping malls, sports complexes, market places, public places like airports, railways stations, and bus stands can help us with the preparedness of repercussion of unusual or violent crowd behaviors. Violence detection aims at identifying whether a violent action has occurred and has evolved as a popular theme in the department of image processing and computer vision. Improved highly effective methods for intelligent analysis are highly demanded. Various methodologies can detect such activities based on Deep Learning algorithms, SVM, and Machine learning algorithms. Deep neural nets and Transfer learning have proven highly successful in the detection of violent activities. The motive of this dissertation is to propose a novel deep ConvNet system for the task of detecting violence by extraction of motion features from RGB Dynamic Motion Images (DMI). Motion feature extraction and prediction of violent content using a stream of RGB DMI is done effectively by pre-trained CNN model – Inception Resnet-V2 followed by fine-tuning layers. The advantages and limitations of existing state-of-the-art CNN based architectures for violence detection suggested by various researchers and popular datasets used for violence detection are also discussed. For performance validation of the proposed novel model, tests are performed on three popular and publically available benchmarks – Hockey Fight dataset, Real Life v Violence Dataset, and movie dataset. The performance is also checked against the other widely used pre-trained models – Resent50 and Inception V3.